NCO 4.6.2 User Guide

This file documents NCO, a collection of utilities to manipulate and analyze netCDF files.

Copyright © 1995–2016 Charlie Zender

This is the first edition of the NCO User Guide,
and is consistent with version 2 of texinfo.tex.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. The license is available online at http://www.gnu.org/copyleft/fdl.html

The original author of this software, Charlie Zender, wants to improve it with the help of your suggestions, improvements, bug-reports, and patches.
Charlie Zender <surname at uci dot edu> (yes, my surname is zender)
3200 Croul Hall
Department of Earth System Science
University of California, Irvine
Irvine, CA 92697-3100

Table of Contents

NCO User Guide

The netCDF Operators, or NCO, are a suite of programs known as operators. The operators facilitate manipulation and analysis of data stored in the self-describing netCDF format, available from (http://www.unidata.ucar.edu/software/netcdf). Each NCO operator (e.g., ncks) takes netCDF input file(s), performs an operation (e.g., averaging, hyperslabbing, or renaming), and outputs a processed netCDF file. Although most users of netCDF data are involved in scientific research, these data formats, and thus NCO, are generic and are equally useful in fields from agriculture to zoology. The NCO User Guide illustrates NCO use with examples from the field of climate modeling and analysis. The NCO homepage is http://nco.sf.net, and the source code is maintained at http://github.com/nco/nco.

This documentation is for NCO version 4.6.2. It was last updated 16 November 2016. Corrections, additions, and rewrites of this documentation are gratefully welcome.

Enjoy,
Charlie Zender

Foreword

NCO is the result of software needs that arose while I worked on projects funded by NCAR, NASA, and ARM. Thinking they might prove useful as tools or templates to others, it is my pleasure to provide them freely to the scientific community. Many users (most of whom I have never met) have encouraged the development of NCO. Thanks espcially to Jan Polcher, Keith Lindsay, Arlindo da Silva, John Sheldon, and William Weibel for stimulating suggestions and correspondence. Your encouragment motivated me to complete the NCO User Guide. So if you like NCO, send me a note! I should mention that NCO is not connected to or officially endorsed by Unidata, ACD, ASP, CGD, or Nike.

Charlie Zender
May 1997
Boulder, Colorado

Major feature improvements entitle me to write another Foreword. In the last five years a lot of work has been done to refine NCO. NCO is now an open source project and appears to be much healthier for it. The list of illustrious institutions that do not endorse NCO continues to grow, and now includes UCI.

Charlie Zender
October 2000
Irvine, California

The most remarkable advances in NCO capabilities in the last few years are due to contributions from the Open Source community. Especially noteworthy are the contributions of Henry Butowsky and Rorik Peterson.

Charlie Zender
January 2003
Irvine, California

NCO was generously supported from 2004–2008 by US National Science Foundation (NSF) grant IIS-0431203. This support allowed me to maintain and extend core NCO code, and others to advance NCO in new directions: Gayathri Venkitachalam helped implement MPI; Harry Mangalam improved regression testing and benchmarking; Daniel Wang developed the server-side capability, SWAMP; and Henry Butowsky, a long-time contributor, developed ncap2. This support also led NCO to debut in professional journals and meetings. The personal and professional contacts made during this evolution have been immensely rewarding.

Charlie Zender
March 2008
Grenoble, France

The end of the NSF SEI grant in August, 2008 curtailed NCO development. Fortunately we could justify supporting Henry Butowsky on other research grants until May, 2010 while he developed the key ncap2 features used in our climate research. And recently the NASA ACCESS program commenced funding us to support netCDF4 group functionality. Thus NCO will grow and evade bit-rot for the foreseeable future.

I continue to receive with gratitude the thanks of NCO users at nearly every scientific meeting I attend. People introduce themselves, shake my hand and extol NCO, often effusively, while I grin in stupid embarassment. These exchanges lighten me like anti-gravity. Sometimes I daydream how many hours NCO has turned from grunt work to productive research for researchers world-wide, or from research into early happy-hours. It’s a cool feeling.

Charlie Zender
April, 2012
Irvine, California

The NASA ACCESS 2011 program generously supported (Cooperative Agreement NNX12AF48A) NCO from 2012–2015. This allowed us to produce the first iteration of a Group-oriented Data Analysis and Distribution (GODAD) software ecosystem. Shifting more geoscience data analysis to GODAD is a long-term plan. Then the NASA ACCESS 2013 program agreed to support (Cooperative Agreement NNX14AH55A) NCO from 2014–2016. This support permits us to implement support for Swath-like Data (SLD). Most recently, the DOE has funded me to implement NCO re-gridding and parallelization in support of their ACME program. After many years of crafting NCO as an after-hours hobby, I finally have the cushion necessary to give it some real attention. And I’m looking forward to this next, and most intense yet, phase of NCO development.

Charlie Zender
June, 2015
Irvine, California

Summary

This manual describes NCO, which stands for netCDF Operators. NCO is a suite of programs known as operators. Each operator is a standalone, command line program executed at the shell-level like, e.g., ls or mkdir. The operators take netCDF files (including HDF5 files constructed using the netCDF API) as input, perform an operation (e.g., averaging or hyperslabbing), and produce a netCDF file as output. The operators are primarily designed to aid manipulation and analysis of data. The examples in this documentation are typical applications of the operators for processing climate model output. This stems from their origin, though the operators are as general as netCDF itself.

1 Introduction

1.1 Availability

The complete NCO source distribution is currently distributed as a compressed tarfile from http://sf.net/projects/nco and from http://dust.ess.uci.edu/nco/nco.tar.gz. The compressed tarfile must be uncompressed and untarred before building NCO. Uncompress the file with ‘gunzip nco.tar.gz’. Extract the source files from the resulting tarfile with ‘tar -xvf nco.tar’. GNU tar lets you perform both operations in one step with ‘tar -xvzf nco.tar.gz’.

The documentation for NCO is called the NCO User Guide. The User Guide is available in PDF, Postscript, HTML, DVI, TeXinfo, and Info formats. These formats are included in the source distribution in the files nco.pdf, nco.ps, nco.html, nco.dvi, nco.texi, and nco.info*, respectively. All the documentation descends from a single source file, nco.texi ¹. Hence the documentation in every format is very similar. However, some of the complex mathematical expressions needed to describe ncwa can only be displayed in DVI, Postscript, and PDF formats.

A complete list of papers and publications on/about NCO is available on the NCO homepage. Most of these are freely available. The primary refereed publications are ZeM06 and Zen08. These contain copyright restrictions which limit their redistribution, but they are freely available in preprint form from the NCO.

If you want to quickly see what the latest improvements in NCO are (without downloading the entire source distribution), visit the NCO homepage at http://nco.sf.net. The HTML version of the User Guide is also available online through the World Wide Web at URL http://nco.sf.net/nco.html. To build and use NCO, you must have netCDF installed. The netCDF homepage is http://www.unidata.ucar.edu/software/netcdf.

New NCO releases are announced on the netCDF list and on the nco-announce mailing list http://lists.sf.net/mailman/listinfo/nco-announce.

1.2 How to Use This Guide

Detailed instructions about how to download the newest version, and how to complie source code, as well as a FAQ and descriptions of Known Problems etc. are on our homepage (http://nco.sf.net/).

There are twelve operators in the current version (4.6.2). The function of each is explained in Reference Manual. Many of the tasks that NCO can accomplish are described during the explanation of common NCO Features (see Shared features). More specific use examples for each operator can be seen by visiting the operator-specific examples in the Reference Manual. These can be found directly by prepending the operator name with the xmp_ tag, e.g., http://nco.sf.net/nco.html#xmp_ncks. Also, users can type the operator name on the shell command line to see all the available options, or type, e.g., ‘man ncks’ to see a help man-page.

NCO is a command-line language. You may either use an operator after the prompt (e.g., ‘$’ here), like,

$ operator [options] input [output]

or write all commands lines into a shell script, as in the CMIP5 Example (see CMIP5 Example).

If you are new to NCO, the Quick Start (see Quick Start) shows simple examples about how to use NCO on different kinds of data files. More detailed “real-world” examples are in the CMIP5 Example. The Index is presents multiple keyword entries for the same subject. If these resources do not help enough, please see Help Requests and Bug Reports.

1.3 Operating systems compatible with NCO

In its time on Earth, NCO has been successfully ported and tested on so many 32- and 64-bit platforms that if we did not write them down here we would forget their names: IBM AIX 4.x, 5.x, FreeBSD 4.x, GNU/Linux 2.x, LinuxPPC, LinuxAlpha, LinuxARM, LinuxSparc64, LinuxAMD64, SGI IRIX 5.x and 6.x, MacOS X 10.x, DEC OSF, NEC Super-UX 10.x, Sun SunOS 4.1.x, Solaris 2.x, Cray UNICOS 8.x–10.x, and Microsoft Windows (95, 98, NT, 2000, XP, Vista, 7, 8, 10). If you port the code to a new operating system, please send me a note and any patches you required.

The major prerequisite for installing NCO on a particular platform is the successful, prior installation of the netCDF library (and, as of 2003, the UDUnits library). Unidata has shown a commitment to maintaining netCDF and UDUnits on all popular UNIX platforms, and is moving towards full support for the Microsoft Windows operating system (OS). Given this, the only difficulty in implementing NCO on a particular platform is standardization of various C-language API system calls. NCO code is tested for ANSI compliance by compiling with C99 compilers including those from GNU (‘gcc -std=c99 -pedantic -D_BSD_SOURCE -D_POSIX_SOURCE’ -Wall) ², Comeau Computing (‘como --c99’), Cray (‘cc’), HP/Compaq/DEC (‘cc’), IBM (‘xlc -c -qlanglvl=extc99’), Intel (‘icc -std=c99’), LLVM (‘clang’), NEC (‘cc’), PathScale (QLogic) (‘pathcc -std=c99’), PGI (‘pgcc -c9x’), SGI (‘cc -c99’), and Sun (‘cc’). NCO (all commands and the libnco library) and the C++ interface to netCDF (called libnco_c++) comply with the ISO C++ standards as implemented by Comeau Computing (‘como’), Cray (‘CC’), GNU (‘g++ -Wall’), HP/Compaq/DEC (‘cxx’), IBM (‘xlC’), Intel (‘icc’), Microsoft (‘MVS’), NEC (‘c++’), PathScale (Qlogic) (‘pathCC’), PGI (‘pgCC’), SGI (‘CC -LANG:std’), and Sun (‘CC -LANG:std’). See nco/bld/Makefile and nco/src/nco_c++/Makefile.old for more details and exact settings.

Until recently (and not even yet), ANSI-compliant has meant compliance with the 1989 ISO C-standard, usually called C89 (with minor revisions made in 1994 and 1995). C89 lacks variable-size arrays, restricted pointers, some useful printf formats, and many mathematical special functions. These are valuable features of C99, the 1999 ISO C-standard. NCO is C99-compliant where possible and C89-compliant where necessary. Certain branches in the code are required to satisfy the native SGI and SunOS C compilers, which are strictly ANSI C89 compliant, and cannot benefit from C99 features. However, C99 features are fully supported by modern AIX, GNU, Intel, NEC, Solaris, and UNICOS compilers. NCO requires a C99-compliant compiler as of NCO version 2.9.8, released in August, 2004.

The most time-intensive portion of NCO execution is spent in arithmetic operations, e.g., multiplication, averaging, subtraction. These operations were performed in Fortran by default until August, 1999. This was a design decision based on the relative speed of Fortran-based object code vs. C-based object code in late 1994. C compiler vectorization capabilities have dramatically improved since 1994. We have accordingly replaced all Fortran subroutines with C functions. This greatly simplifies the task of building NCO on nominally unsupported platforms. As of August 1999, NCO built entirely in C by default. This allowed NCO to compile on any machine with an ANSI C compiler. In August 2004, the first C99 feature, the restrict type qualifier, entered NCO in version 2.9.8. C compilers can obtain better performance with C99 restricted pointers since they inform the compiler when it may make Fortran-like assumptions regarding pointer contents alteration. Subsequently, NCO requires a C99 compiler to build correctly ³.

In January 2009, NCO version 3.9.6 was the first to link to the GNU Scientific Library (GSL). GSL must be version 1.4 or later. NCO, in particular ncap2, uses the GSL special function library to evaluate geoscience-relevant mathematics such as Bessel functions, Legendre polynomials, and incomplete gamma functions (see GSL special functions).

In June 2005, NCO version 3.0.1 began to take advantage of C99 mathematical special functions. These include the standarized gamma function (called tgamma() for “true gamma”). NCO automagically takes advantage of some GNU Compiler Collection (GCC) extensions to ANSI C.

As of July 2000 and NCO version 1.2, NCO no longer performs arithmetic operations in Fortran. We decided to sacrifice executable speed for code maintainability. Since no objective statistics were ever performed to quantify the difference in speed between the Fortran and C code, the performance penalty incurred by this decision is unknown. Supporting Fortran involves maintaining two sets of routines for every arithmetic operation. The USE_FORTRAN_ARITHMETIC flag is still retained in the Makefile. The file containing the Fortran code, nco_fortran.F, has been deprecated but a volunteer (Dr. Frankenstein?) could resurrect it. If you would like to volunteer to maintain nco_fortran.F please contact me.

1.3.1 Compiling NCO for Microsoft Windows OS

NCO has been successfully ported and tested on most Microsoft Windows operating systems including: XP SP2/Vista/7. Support is provided for compiling either native Windows executables, using the Microsoft Visual Studio 2010 Compiler, or with Cygwin, the UNIX-emulating compatibility layer with the GNU toolchain. The switches necessary to accomplish both are included in the standard distribution of NCO.

Using Microsoft Visual Studio (MVS), one must build NCO with the C++ compiler since MVS does not support C99. Qt, a convenient integrated development environment, was used to convert the project files to MVS format. The Qt files themselves are distributed in the nco/qt directory.

Using the freely available Cygwin (formerly gnu-win32) development environment ⁴, the compilation process is very similar to installing NCO on a UNIX system. Set the PVM_ARCH preprocessor token to WIN32. Note that defining WIN32 has the side effect of disabling Internet features of NCO (see below). NCO should now build like it does on UNIX.

The least portable section of the code is the use of standard UNIX and Internet protocols (e.g., ftp, rcp, scp, sftp, getuid, gethostname, and header files <arpa/nameser.h> and <resolv.h>). Fortunately, these UNIX-y calls are only invoked by the single NCO subroutine which is responsible for retrieving files stored on remote systems (see Remote storage). In order to support NCO on the Microsoft Windows platforms, this single feature was disabled (on Windows OS only). This was required by Cygwin 18.x—newer versions of Cygwin may support these protocols (let me know if this is the case). The NCO operators should behave identically on Windows and UNIX platforms in all other respects.

1.4 Symbolic Links

NCO relies on a common set of underlying algorithms. To minimize duplication of source code, multiple operators sometimes share the same underlying source. This is accomplished by symbolic links from a single underlying executable program to one or more invoked executable names. For example, nces and ncrcat are symbolically linked to the ncra executable. The ncra executable behaves slightly differently based on its invocation name (i.e., ‘argv[0]’), which can be nces, ncra, or ncrcat. Logically, these are three different operators that happen to share the same executable.

For historical reasons, and to be more user friendly, multiple synonyms (or pseudonyms) may refer to the same operator invoked with different switches. For example, ncdiff is the same as ncbo and ncpack is the same as ncpdq. We implement the symbolic links and synonyms by the executing the following UNIX commands in the directory where the NCO executables are installed.

ln -s -f ncbo ncdiff    # ncbo --op_typ='-'
ln -s -f ncra nces      # ncra --pseudonym='nces'
ln -s -f ncra ncrcat    # ncra --pseudonym='ncrcat'
ln -s -f ncbo ncadd     # ncbo --op_typ='+'
ln -s -f ncbo ncsubtract # ncbo --op_typ='-'
ln -s -f ncbo ncmultiply # ncbo --op_typ='*'
ln -s -f ncbo ncdivide   # ncbo --op_typ='/'
ln -s -f ncpdq ncpack    # ncpdq
ln -s -f ncpdq ncunpack  # ncpdq --unpack
# NB: Windows/Cygwin executable/link names have '.exe' suffix, e.g.,
ln -s -f ncbo.exe ncdiff.exe
...

The imputed command called by the link is given after the comment. As can be seen, some these links impute the passing of a command line argument to further modify the behavior of the underlying executable. For example, ncdivide is a pseudonym for ncbo --op_typ='/'.

1.5 Libraries

Like all executables, the NCO operators can be built using dynamic linking. This reduces the size of the executable and can result in significant performance enhancements on multiuser systems. Unfortunately, if your library search path (usually the LD_LIBRARY_PATH environment variable) is not set correctly, or if the system libraries have been moved, renamed, or deleted since NCO was installed, it is possible NCO operators will fail with a message that they cannot find a dynamically loaded (aka shared object or ‘.so’) library. This will produce a distinctive error message, such as ‘ld.so.1: /usr/local/bin/nces: fatal: libsunmath.so.1: can't open file: errno=2’. If you received an error message like this, ask your system administrator to diagnose whether the library is truly missing ⁵, or whether you simply need to alter your library search path. As a final remedy, you may re-compile and install NCO with all operators statically linked.

1.6 netCDF2/3/4 and HDF4/5 Support

netCDF version 2 was released in 1993. NCO (specifically ncks) began soon after this in 1994. netCDF 3.0 was released in 1996, and we were not exactly eager to convert all code to the newer, less tested netCDF implementation. One netCDF3 interface call (nc_inq_libvers) was added to NCO in January, 1998, to aid in maintainance and debugging. In March, 2001, the final NCO conversion to netCDF3 was completed (coincidentally on the same day netCDF 3.5 was released). NCO versions 2.0 and higher are built with the -DNO_NETCDF_2 flag to ensure no netCDF2 interface calls are used.

However, the ability to compile NCO with only netCDF2 calls is worth maintaining because HDF version 4, aka HDF4 or simply HDF, ⁶ (available from HDF) supports only the netCDF2 library calls (see http://hdfgroup.org/UG41r3_html/SDS_SD.fm12.html#47784). There are two versions of HDF. Currently HDF version 4.x supports the full netCDF2 API and thus NCO version 1.2.x. If NCO version 1.2.x (or earlier) is built with only netCDF2 calls then all NCO operators should work with HDF4 files as well as netCDF files ⁷. The preprocessor token NETCDF2_ONLY exists in NCO version 1.2.x to eliminate all netCDF3 calls. Only versions of NCO numbered 1.2.x and earlier have this capability.

HDF version 5 became available in 1999, but did not support netCDF (or, for that matter, Fortran) as of December 1999. By early 2001, HDF5 did support Fortran90. Thanks to an NSF-funded “harmonization” partnership, HDF began to fully support the netCDF3 read interface (which is employed by NCO 2.x and later). In 2004, Unidata and THG began a project to implement the HDF5 features necessary to support the netCDF API. NCO version 3.0.3 added support for reading/writing netCDF4-formatted HDF5 files in October, 2005. See File Formats and Conversion for more details.

HDF support for netCDF was completed with HDF5 version version 1.8 in 2007. The netCDF front-end that uses this HDF5 back-end was completed and released soon after as netCDF version 4. Download it from the netCDF4 website.

NCO version 3.9.0, released in May, 2007, added support for all netCDF4 atomic data types except NC_STRING. Support for NC_STRING, including ragged arrays of strings, was finally added in version 3.9.9, released in June, 2009. Support for additional netCDF4 features has been incremental. We add one netCDF4 feature at a time. You must build NCO with netCDF4 to obtain this support.

NCO supports many netCDF4 features including atomic data types, Lempel-Ziv compression (deflation), chunking, and groups. The new atomic data types are NC_UBYTE, NC_USHORT, NC_UINT, NC_INT64, and NC_UINT64. Eight-byte integer support is an especially useful improvement from netCDF3. All NCO operators support these types, e.g., ncks copies and prints them, ncra averages them, and ncap2 processes algebraic scripts with them. ncks prints compression information, if any, to screen.

NCO version 3.9.1 (June, 2007) added support for netCDF4 Lempel-Ziv deflation. Lempel-Ziv deflation is a lossless compression technique. See Deflation for more details.

NCO version 3.9.9 (June, 2009) added support for netCDF4 chunking in ncks and ncecat. NCO version 4.0.4 (September, 2010) completed support for netCDF4 chunking in the remaining operators. See Chunking for more details.

NCO version 4.2.2 (October, 2012) added support for netCDF4 groups in ncks and ncecat. Group support for these operators was complete (e.g., regular expressions to select groups and Group Path Editing) as of NCO version 4.2.6 (March, 2013). See Group Path Editing for more details. Group support for all other operators was finished in the NCO version 4.3.x series completed in December, 2013.

Support for netCDF4 in the first arithmetic operator, ncbo, was introduced in NCO version 4.3.0 (March, 2013). NCO version 4.3.1 (May, 2013) completed this support and introduced the first example of automatic group broadcasting. See ncbo netCDF Binary Operator for more details.

netCDF4-enabled NCO handles netCDF3 files without change. In addition, it automagically handles netCDF4 (HDF5) files: If you feed NCO netCDF3 files, it produces netCDF3 output. If you feed NCO netCDF4 files, it produces netCDF4 output. Use the handy-dandy ‘-4’ switch to request netCDF4 output from netCDF3 input, i.e., to convert netCDF3 to netCDF4. See File Formats and Conversion for more details.

When linked to a netCDF library that was built with HDF4 support ⁸, NCO automatically supports reading HDF4 files and writing them as netCDF3/netCDF4/HDF5 files. NCO can only write through the netCDF API, which can only write netCDF3/netCDF4/HDF5 files. So NCO can read HDF4 files, perform manipulations and calculations, and then it must write the results in netCDF format.

NCO support for HDF4 has been quite functional since December, 2013. For best results install NCO versions 4.4.0 or later on top of netCDF versions 4.3.1 or later. Getting to this point has been an iterative effort where Unidata improved netCDF library capabilities in response to our requests. NCO versions 4.3.6 and earlier do not explicitly support HDF4, yet should work with HDF4 if compiled with a version of netCDF (4.3.2 or later?) that does not unexpectedly die when probing HDF4 files with standard netCDF calls. NCO versions 4.3.7–4.3.9 (October–December, 2013) use a special flag to workaround netCDF HDF4 issues. The user must tell these versions of NCO that an input file is HDF4 format by using the ‘--hdf4’ switch.

When compiled with netCDF version 4.3.1 (20140116) or later, NCO versions 4.4.0 (January, 2014) and later more gracefully handle HDF4 files. In particular, the ‘--hdf4’ switch is obsolete. Current versions of NCO use netCDF to determine automatically whether the underlying file is HDF4, and then take appropriate precautions to avoid netCDF4 API calls that fail when applied to HDF4 files (e.g., nc_inq_var_chunking(), nc_inq_var_deflate()). When compiled with netCDF version 4.3.2 (20140423) or earlier, NCO will report that chunking and deflation properties of HDF4 files as HDF4_UNKNOWN, because determining those properties was impossible. When compiled with netCDF version 4.3.3-rc2 (20140925) or later, NCO versions 4.4.6 (October, 2014) and later fully support chunking and deflation features of HDF4 files. The ‘--hdf4’ switch is supported (for backwards compatibility) yet redundant (i.e., does no harm) with current versions of NCO and netCDF.

Converting HDF4 files to netCDF: Since NCO reads HDF4 files natively, it is now easy to convert HDF4 files to netCDF files directly, e.g.,

ncks        fl.hdf fl.nc # Convert HDF4->netCDF4 (NCO 4.4.0+, netCDF 4.3.1+)
ncks --hdf4 fl.hdf fl.nc # Convert HDF4->netCDF4 (NCO 4.3.7-4.3.9)

The most efficient and accurate way to convert HDF4 data to netCDF format is to convert to netCDF4 using NCO as above. Many HDF4 producers (NASA!) love to use netCDF4 types, e.g., unsigned bytes, so this procedure is the most typical. Conversion of HDF4 to netCDF4 as above suffices when the data will only be processed by NCO and other netCDF4-aware tools.

However, many tools are not fully netCDF4-aware, and so conversion to netCDF3 may be desirable. Obtaining any netCDF file from an HDF4 is easy:

ncks -3 fl.hdf fl.nc      # HDF4->netCDF3 (NCO 4.4.0+, netCDF 4.3.1+)
ncks -4 fl.hdf fl.nc      # HDF4->netCDF4 (NCO 4.4.0+, netCDF 4.3.1+)
ncks -6 fl.hdf fl.nc      # HDF4->netCDF3 64-bit  (NCO 4.4.0+, ...)
ncks -7 -L 1 fl.hdf fl.nc # HDF4->netCDF4 classic (NCO 4.4.0+, ...)
ncks --hdf4 -3 fl.hdf fl.nc # HDF4->netCDF3 (netCDF 4.3.0-)
ncks --hdf4 -4 fl.hdf fl.nc # HDF4->netCDF4 (netCDF 4.3.0-)
ncks --hdf4 -6 fl.hdf fl.nc # HDF4->netCDF3 64-bit  (netCDF 4.3.0-)
ncks --hdf4 -7 fl.hdf fl.nc # HDF4->netCDF4 classic (netCDF 4.3.0-)

As of NCO version 4.4.0 (January, 2014), these commands work even when the HDF4 file contains netCDF4 atomic types (e.g., unsigned bytes, 64-bit integers) because NCO can autoconvert everything to atomic types supported by netCDF3 ⁹.

As of NCO version 4.4.4 (May, 2014) both ncl_convert2nc and NCO have built-in, automatic workarounds to handle element names that contain characters that are legal in HDF though are illegal in netCDF. For example, slashes and leading special characters are are legal in HDF and illegal in netCDF element (i.e., group, variable, dimension, and attribute) names. NCO converts these forbidden characters to underscores, and retains the original names of variables in automatically produced attributes named hdf_name ¹⁰.

Finally, in February 2014, we learned that the HDF group has a project called H4CF (described here) whose goal is to make HDF4 files accessible to CF tools and conventions. Their project includes a tool named h4tonccf that converts HDF4 files to netCDF3 or netCDF4 files. We are not yet sure what advantages or features h4tonccf has that are not in NCO, though we suspect both methods have their own advantages. Corrections welcome.

As of 2012, netCDF4 is relatively stable software. Problems with netCDF4 and HDF libraries have mainly been fixed. Binary NCO distributions shipped as RPMs and as debs have used the netCDF4 library since 2010 and 2011, respectively.

One must often build NCO from source to obtain netCDF4 support. Typically, one specifies the root of the netCDF4 installation directory. Do this with the NETCDF4_ROOT variable. Then use your preferred NCO build mechanism, e.g.,

export NETCDF4_ROOT=/usr/local/netcdf4 # Set netCDF4 location
cd ~/nco;./configure --enable-netcdf4  # Configure mechanism -or-
cd ~/nco/bld;./make NETCDF4=Y allinone # Old Makefile mechanism

We carefully track the netCDF4 releases, and keep the netCDF4 atomic type support and other features working. Our long term goal is to utilize more of the extensive new netCDF4 feature set. The next major netCDF4 feature we are likely to utilize is parallel I/O. We will enable this in the MPI netCDF operators.

1.7 Help Requests and Bug Reports

We generally receive three categories of mail from users: help requests, bug reports, and feature requests. Notes saying the equivalent of “Hey, NCO continues to work great and it saves me more time everyday than it took to write this note” are a distant fourth.

There is a different protocol for each type of request. The preferred etiquette for all communications is via NCO Project Forums. Do not contact project members via personal e-mail unless your request comes with money or you have damaging information about our personal lives. Please use the Forums—they preserve a record of the questions and answers so that others can learn from our exchange. Also, since NCO is government-funded, this record helps us provide program officers with information they need to evaluate our project.

Before posting to the NCO forums described below, you might first register your name and email address with SourceForge.net or else all of your postings will be attributed to nobody. Once registered you may choose to monitor any forum and to receive (or not) email when there are any postings including responses to your questions. We usually reply to the forum message, not to the original poster.

If you want us to include a new feature in NCO, check first to see if that feature is already on the TODO list. If it is, why not implement that feature yourself and send us the patch? If the feature is not yet on the list, then send a note to the NCO Discussion forum.

Read the manual before reporting a bug or posting a help request. Sending questions whose answers are not in the manual is the best way to motivate us to write more documentation. We would also like to accentuate the contrapositive of this statement. If you think you have found a real bug the most helpful thing you can do is simplify the problem to a manageable size and then report it. The first thing to do is to make sure you are running the latest publicly released version of NCO.

Once you have read the manual, if you are still unable to get NCO to perform a documented function, submit a help request. Follow the same procedure as described below for reporting bugs (after all, it might be a bug). That is, describe what you are trying to do, and include the complete commands (run with ‘-D 5’), error messages, and version of NCO (with ‘-r’). Post your help request to the NCO Help forum.

If you think you used the right command when NCO misbehaves, then you might have found a bug. Incorrect numerical answers are the highest priority. We usually fix those within one or two days. Core dumps and sementation violations receive lower priority. They are always fixed, eventually.

How do you simplify a problem that reveal a bug? Cut out extraneous variables, dimensions, and metadata from the offending files and re-run the command until it no longer breaks. Then back up one step and report the problem. Usually the file(s) will be very small, i.e., one variable with one or two small dimensions ought to suffice. Run the operator with ‘-r’ and then run the command with ‘-D 5’ to increase the verbosity of the debugging output. It is very important that your report contain the exact error messages and compile-time environment. Include a copy of your sample input file, or place one on a publicly accessible location, of the file(s). If you are sure it is a bug, post the full report to the NCO Project buglist. Otherwise post all the information to NCO Help forum.

Build failures count as bugs. Our limited machine access means we cannot fix all build failures. The information we need to diagnose, and often fix, build failures are the three files output by GNU build tools, nco.config.log.${GNU_TRP}.foo, nco.configure.${GNU_TRP}.foo, and nco.make.${GNU_TRP}.foo. The file configure.eg shows how to produce these files. Here ${GNU_TRP} is the “GNU architecture triplet”, the chip-vendor-OS string returned by config.guess. Please send us your improvements to the examples supplied in configure.eg. The regressions archive at http://dust.ess.uci.edu/nco/rgr contains the build output from our standard test systems. You may find you can solve the build problem yourself by examining the differences between these files and your own.

2 Operator Strategies

2.1 Philosophy

The main design goal is command line operators which perform useful, scriptable operations on netCDF files. Many scientists work with models and observations which produce too much data to analyze in tabular format. Thus, it is often natural to reduce and massage this raw or primary level data into summary, or second level data, e.g., temporal or spatial averages. These second level data may become the inputs to graphical and statistical packages, and are often more suitable for archival and dissemination to the scientific community. NCO performs a suite of operations useful in manipulating data from the primary to the second level state. Higher level interpretive languages (e.g., IDL, Yorick, Matlab, NCL, Perl, Python), and lower level compiled languages (e.g., C, Fortran) can always perform any task performed by NCO, but often with more overhead. NCO, on the other hand, is limited to a much smaller set of arithmetic and metadata operations than these full blown languages.

Another goal has been to implement enough command line switches so that frequently used sequences of these operators can be executed from a shell script or batch file. Finally, NCO was written to consume the absolute minimum amount of system memory required to perform a given job. The arithmetic operators are extremely efficient; their exact memory usage is detailed in Memory Requirements.

2.2 Climate Model Paradigm

NCO was developed at NCAR to aid analysis and manipulation of datasets produced by General Circulation Models (GCMs). GCM datasets share many features with other gridded scientific datasets and so provide a useful paradigm for the explication of the NCO operator set. Examples in this manual use a GCM paradigm because latitude, longitude, time, temperature and other fields related to our natural environment are as easy to visualize for the layman as the expert.

2.3 Temporary Output Files

NCO operators are designed to be reasonably fault tolerant, so that a system failure or user-abort of the operation (e.g., with C-c) does not cause loss of data. The user-specified output-file is only created upon successful completion of the operation ¹¹. This is accomplished by performing all operations in a temporary copy of output-file. The name of the temporary output file is constructed by appending .pid<process ID>.<operator name>.tmp to the user-specified output-file name. When the operator completes its task with no fatal errors, the temporary output file is moved to the user-specified output-file. This imbues the process with fault-tolerance since fatal error (e.g., disk space fills up) affect only the temporary output file, leaving the final output file not created if it did not already exist. Note the construction of a temporary output file uses more disk space than just overwriting existing files “in place” (because there may be two copies of the same file on disk until the NCO operation successfully concludes and the temporary output file overwrites the existing output-file). Also, note this feature increases the execution time of the operator by approximately the time it takes to copy the output-file ¹². Finally, note this fault-tolerant feature allows the output-file to be the same as the input-file without any danger of “overlap”.

Over time many “power users” have requested a way to turn-off the fault-tolerance safety feature of automatically creating a temporary file. Often these users build and execute production data analysis scripts that are repeated frequently on large datasets. Obviating an extra file write can then conserve significant disk space and time. For this purpose NCO has, since version 4.2.1 in August, 2012, made configurable the controls over temporary file creation. The ‘--wrt_tmp_fl’ and equivalent ‘--write_tmp_fl’ switches ensure NCO writes output to an intermediate temporary file. This is and has always been the default behavior so there is currently no need to specify these switches. However, the default may change some day, especially since writing to RAM disks (see RAM disks) may some day become the default. The ‘--no_tmp_fl’ switch causes NCO to write directly to the final output file instead of to an intermediate temporary file. “Power users” may wish to invoke this switch to increase performance (i.e., reduce wallclock time) when manipulating large files. When eschewing temporary files, users may forsake the ability to have the same name for both output-file and input-file since, as described above, the temporary file prevented overlap issues. However, if the user creates the output file in RAM (see RAM disks) then it is still possible to have the same name for both output-file and input-file.

ncks in.nc out.nc # Default: create out.pid.tmp.nc then move to out.nc
ncks --wrt_tmp_fl in.nc out.nc # Same as default
ncks --no_tmp_fl in.nc out.nc # Create out.nc directly on disk
ncks --no_tmp_fl in.nc in.nc # ERROR-prone! Overwrite in.nc with itself
ncks --create_ram --no_tmp_fl in.nc in.nc # Create in RAM, write to disk
ncks --open_ram --no_tmp_fl in.nc in.nc # Read into RAM, write to disk

There is no reason to expect the fourth example to work. The behavior of overwriting a file while reading from the same file is undefined, much as is the shell command ‘cat foo > foo’. Although it may “work” in some cases, it is unreliable. One way around this is to use ‘--create_ram’ so that the output file is not written to disk until the input file is closed, See RAM disks. However, as of 20130328, the behavior of the ‘--create_ram’ and ‘--open_ram’ examples has not been thoroughly tested.

The NCO authors have seen compelling use cases for utilizing the RAM switches, though not (yet) for combining them with ‘--no_tmp_fl’. NCO implements both options because they are largely independent of eachother. It is up to “power users” to discover which best fit their needs. We welcome accounts of your experiences posted to the forums.

Other safeguards exist to protect the user from inadvertently overwriting data. If the output-file specified for a command is a pre-existing file, then the operator will prompt the user whether to overwrite (erase) the existing output-file, attempt to append to it, or abort the operation. However, in processing large amounts of data, too many interactive questions slows productivity. Therefore NCO also implements two ways to override its own safety features, the ‘-O’ and ‘-A’ switches. Specifying ‘-O’ tells the operator to overwrite any existing output-file without prompting the user interactively. Specifying ‘-A’ tells the operator to attempt to append to any existing output-file without prompting the user interactively. These switches are useful in batch environments because they suppress interactive keyboard input.

2.4 Appending Variables

Adding variables from one file to another is often desirable. This is referred to as appending, although some prefer the terminology merging ¹³ or pasting. Appending is often confused with what NCO calls concatenation. In NCO, concatenation refers to splicing a variable along the record dimension. The length along the record dimension of the output is the sum of the lengths of the input files. Appending, on the other hand, refers to copying a variable from one file to another file which may or may not already contain the variable ¹⁴. NCO can append or concatenate just one variable, or all the variables in a file at the same time.

In this sense, ncks can append variables from one file to another file. This capability is invoked by naming two files on the command line, input-file and output-file. When output-file already exists, the user is prompted whether to overwrite, append/replace, or exit from the command. Selecting overwrite tells the operator to erase the existing output-file and replace it with the results of the operation. Selecting exit causes the operator to exit—the output-file will not be touched in this case. Selecting append/replace causes the operator to attempt to place the results of the operation in the existing output-file, See ncks netCDF Kitchen Sink.

The simplest way to create the union of two files is

ncks -A fl_1.nc fl_2.nc

This puts the contents of fl_1.nc into fl_2.nc. The ‘-A’ is optional. On output, fl_2.nc is the union of the input files, regardless of whether they share dimensions and variables, or are completely disjoint. The append fails if the input files have differently named record dimensions (since netCDF supports only one), or have dimensions of the same name but different sizes.

2.5 Simple Arithmetic and Interpolation

Users comfortable with NCO semantics may find it easier to perform some simple mathematical operations in NCO rather than higher level languages. ncbo (see ncbo netCDF Binary Operator) does file addition, subtraction, multiplication, division, and broadcasting. It even does group broadcasting. ncflint (see ncflint netCDF File Interpolator) does file addition, subtraction, multiplication and interpolation. Sequences of these commands can accomplish simple yet powerful operations from the command line.

2.6 Statistics vs. Concatenation

The most frequently used operators of NCO are probably the statisticians (i.e., tools that do statistics) and concatenators. Because there are so many types of statistics like averaging (e.g., across files, within a file, over the record dimension, over other dimensions, with or without weights and masks) and of concatenating (across files, along the record dimension, along other dimensions), there are currently no fewer than five operators which tackle these two purposes: ncra, nces, ncwa, ncrcat, and ncecat. These operators do share many capabilities ¹⁵, though each has its unique specialty. Two of these operators, ncrcat and ncecat, concatenate hyperslabs across files. The other two operators, ncra and nces, compute statistics across (and/or within) files ¹⁶. First, let’s describe the concatenators, then the statistics tools.

2.6.1 Concatenators ncrcat and ncecat

Joining together independent files along a common record dimension is called concatenation. ncrcat is designed for concatenating record variables, while ncecat is designed for concatenating fixed length variables. Consider five files, 85.nc, 86.nc, … 89.nc each containing a year’s worth of data. Say you wish to create from them a single file, 8589.nc containing all the data, i.e., spanning all five years. If the annual files make use of the same record variable, then ncrcat will do the job nicely with, e.g., ncrcat 8?.nc 8589.nc. The number of records in the input files is arbitrary and can vary from file to file. See ncrcat netCDF Record Concatenator, for a complete description of ncrcat.

However, suppose the annual files have no record variable, and thus their data are all fixed length. For example, the files may not be conceptually sequential, but rather members of the same group, or ensemble. Members of an ensemble may have no reason to contain a record dimension. ncecat will create a new record dimension (named record by default) with which to glue together the individual files into the single ensemble file. If ncecat is used on files which contain an existing record dimension, that record dimension is converted to a fixed-length dimension of the same name and a new record dimension (named record) is created. Consider five realizations, 85a.nc, 85b.nc, … 85e.nc of 1985 predictions from the same climate model. Then ncecat 85?.nc 85_ens.nc glues together the individual realizations into the single file, 85_ens.nc. If an input variable was dimensioned [lat,lon], it will have dimensions [record,lat,lon] in the output file. A restriction of ncecat is that the hyperslabs of the processed variables must be the same from file to file. Normally this means all the input files are the same size, and contain data on different realizations of the same variables. See ncecat netCDF Ensemble Concatenator, for a complete description of ncecat.

ncpdq makes it possible to concatenate files along any dimension, not just the record dimension. First, use ncpdq to convert the dimension to be concatenated (i.e., extended with data from other files) into the record dimension. Second, use ncrcat to concatenate these files. Finally, if desirable, use ncpdq to revert to the original dimensionality. As a concrete example, say that files x_01.nc, x_02.nc, … x_10.nc contain time-evolving datasets from spatially adjacent regions. The time and spatial coordinates are time and x, respectively. Initially the record dimension is time. Our goal is to create a single file that contains joins all the spatially adjacent regions into one single time-evolving dataset.

for idx in 01 02 03 04 05 06 07 08 09 10; do # Bourne Shell
  ncpdq -a x,time x_${idx}.nc foo_${idx}.nc  # Make x record dimension
done
ncrcat foo_??.nc out.nc       # Concatenate along x
ncpdq -a time,x out.nc out.nc # Revert to time as record dimension

Note that ncrcat will not concatenate fixed-length variables, whereas ncecat concatenates both fixed-length and record variables along a new record variable. To conserve system memory, use ncrcat where possible.

2.6.2 Averagers nces, ncra, and ncwa

The differences between the averagers ncra and nces are analogous to the differences between the concatenators. ncra is designed for averaging record variables from at least one file, while nces is designed for averaging fixed length variables from multiple files. ncra performs a simple arithmetic average over the record dimension of all the input files, with each record having an equal weight in the average. nces performs a simple arithmetic average of all the input files, with each file having an equal weight in the average. Note that ncra cannot average fixed-length variables, but nces can average both fixed-length and record variables. To conserve system memory, use ncra rather than nces where possible (e.g., if each input-file is one record long). The file output from nces will have the same dimensions (meaning dimension names as well as sizes) as the input hyperslabs (see nces netCDF Ensemble Statistics, for a complete description of nces). The file output from ncra will have the same dimensions as the input hyperslabs except for the record dimension, which will have a size of 1 (see ncra netCDF Record Averager, for a complete description of ncra).

2.6.3 Interpolator ncflint

ncflint can interpolate data between or two files. Since no other operators have this ability, the description of interpolation is given fully on the ncflint reference page (see ncflint netCDF File Interpolator). Note that this capability also allows ncflint to linearly rescale any data in a netCDF file, e.g., to convert between differing units.

2.7 Large Numbers of Files

Occasionally one desires to digest (i.e., concatenate or average) hundreds or thousands of input files. Unfortunately, data archives (e.g., NASA EOSDIS) may not name netCDF files in a format understood by the ‘-n loop’ switch (see Specifying Input Files) that automagically generates arbitrary numbers of input filenames. The ‘-n loop’ switch has the virtue of being concise, and of minimizing the command line. This helps keeps output file small since the command line is stored as metadata in the history attribute (see History Attribute). However, the ‘-n loop’ switch is useless when there is no simple, arithmetic pattern to the input filenames (e.g., h00001.nc, h00002.nc, … h90210.nc). Moreover, filename globbing does not work when the input files are too numerous or their names are too lengthy (when strung together as a single argument) to be passed by the calling shell to the NCO operator ¹⁷. When this occurs, the ANSI C-standard argc-argv method of passing arguments from the calling shell to a C-program (i.e., an NCO operator) breaks down. There are (at least) three alternative methods of specifying the input filenames to NCO in environment-limited situations.

The recommended method for sending very large numbers (hundreds or more, typically) of input filenames to the multi-file operators is to pass the filenames with the UNIX standard input feature, aka stdin:

# Pipe large numbers of filenames to stdin
/bin/ls | grep ${CASEID}_'......'.nc | ncecat -o foo.nc

This method avoids all constraints on command line size imposed by the operating system. A drawback to this method is that the history attribute (see History Attribute) does not record the name of any input files since the names were not passed on the command line. This makes determining the data provenance at a later date difficult. To remedy this situation, multi-file operators store the number of input files in the nco_input_file_number global attribute and the input file list itself in the nco_input_file_list global attribute (see File List Attributes). Although this does not preserve the exact command used to generate the file, it does retains all the information required to reconstruct the command and determine the data provenance.

A second option is to use the UNIX xargs command. This simple example selects as input to xargs all the filenames in the current directory that match a given pattern. For illustration, consider a user trying to average millions of files which each have a six character filename. If the shell buffer cannot hold the results of the corresponding globbing operator, ??????.nc, then the filename globbing technique will fail. Instead we express the filename pattern as an extended regular expression, ......\.nc (see Subsetting Files). We use grep to filter the directory listing for this pattern and to pipe the results to xargs which, in turn, passes the matching filenames to an NCO multi-file operator, e.g., ncecat.

# Use xargs to transfer filenames on the command line
/bin/ls | grep ${CASEID}_'......'.nc | xargs -x ncecat -o foo.nc

The single quotes protect the only sensitive parts of the extended regular expression (the grep argument), and allow shell interpolation (the ${CASEID} variable substitution) to proceed unhindered on the rest of the command. xargs uses the UNIX pipe feature to append the suitably filtered input file list to the end of the ncecat command options. The -o foo.nc switch ensures that the input files supplied by xargs are not confused with the output file name. xargs does, unfortunately, have its own limit (usually about 20,000 characters) on the size of command lines it can pass. Give xargs the ‘-x’ switch to ensure it dies if it reaches this internal limit. When this occurs, use either the stdin method above, or the symbolic link presented next.

Even when its internal limits have not been reached, the xargs technique may not be sophisticated enough to handle all situations. A full scripting language like Perl can handle any level of complexity of filtering input filenames, and any number of filenames. The technique of last resort is to write a script that creates symbolic links between the irregular input filenames and a set of regular, arithmetic filenames that the ‘-n loop’ switch understands. For example, the following Perl script creates a monotonically enumerated symbolic link to up to one million .nc files in a directory. If there are 999,999 netCDF files present, the links are named 000001.nc to 999999.nc:

# Create enumerated symbolic links
/bin/ls | grep \.nc | perl -e \
'$idx=1;while(<STDIN>){chop;symlink $_,sprintf("%06d.nc",$idx++);}'
ncecat -n 999999,6,1 000001.nc foo.nc
# Remove symbolic links when finished
/bin/rm ??????.nc

The ‘-n loop’ option tells the NCO operator to automatically generate the filnames of the symbolic links. This circumvents any OS and shell limits on command line size. The symbolic links are easily removed once NCO is finished. One drawback to this method is that the history attribute (see History Attribute) retains the filename list of the symbolic links, rather than the data files themselves. This makes it difficult to determine the data provenance at a later date.

2.8 Large Datasets

Large datasets are those files that are comparable in size to the amount of random access memory (RAM) in your computer. Many users of NCO work with files larger than 100 MB. Files this large not only push the current edge of storage technology, they present special problems for programs which attempt to access the entire file at once, such as nces and ncecat. If you work with a 300 MB files on a machine with only 32 MB of memory then you will need large amounts of swap space (virtual memory on disk) and NCO will work slowly, or even fail. There is no easy solution for this. The best strategy is to work on a machine with sufficient amounts of memory and swap space. Since about 2004, many users have begun to produce or analyze files exceeding 2 GB in size. These users should familiarize themselves with NCO’s Large File Support (LFS) capabilities (see Large File Support). The next section will increase your familiarity with NCO’s memory requirements. With this knowledge you may re-design your data reduction approach to divide the problem into pieces solvable in memory-limited situations.

If your local machine has problems working with large files, try running NCO from a more powerful machine, such as a network server. If you get a memory-related core dump (e.g., ‘Error exit (core dumped)’) on a GNU/Linux system, or the operation ends before the entire output file is written, try increasing the process-available memory with ulimit:

ulimit -f unlimited

This may solve constraints on clusters where sufficient hardware resources exist yet where system administrators felt it wise to prevent any individual user from consuming too much of resource. Certain machine architectures, e.g., Cray UNICOS, have special commands which allow one to increase the amount of interactive memory. On Cray systems, try to increase the available memory with the ilimit command.

The speed of the NCO operators also depends on file size. When processing large files the operators may appear to hang, or do nothing, for large periods of time. In order to see what the operator is actually doing, it is useful to activate a more verbose output mode. This is accomplished by supplying a number greater than 0 to the ‘-D debug-level’ (or ‘--debug-level’, or ‘--dbg_lvl’) switch. When the debug-level is nonzero, the operators report their current status to the terminal through the stderr facility. Using ‘-D’ does not slow the operators down. Choose a debug-level between 1 and 3 for most situations, e.g., nces -D 2 85.nc 86.nc 8586.nc. A full description of how to estimate the actual amount of memory the multi-file NCO operators consume is given in Memory Requirements.

2.9 Memory Requirements

Many people use NCO on gargantuan files which dwarf the memory available (free RAM plus swap space) even on today’s powerful machines. These users want NCO to consume the least memory possible so that their scripts do not have to tediously cut files into smaller pieces that fit into memory. We commend these greedy users for pushing NCO to its limits!

This section describes the memory NCO requires during operation. The required memory is based on the underlying algorithms. The description below is the memory usage per thread. Users with shared memory machines may use the threaded NCO operators (see OpenMP Threading). The peak and sustained memory usage will scale accordingly, i.e., by the number of threads. Memory consumption patterns of all operators are similar, with the exception of ncap2.

2.9.1 Single and Multi-file Operators

The multi-file operators currently comprise the record operators, ncra and ncrcat, and the ensemble operators, nces and ncecat. The record operators require much less memory than the ensemble operators. This is because the record operators operate on one single record (i.e., time-slice) at a time, whereas the ensemble operators retrieve the entire variable into memory. Let MS be the peak sustained memory demand of an operator, FT be the memory required to store the entire contents of all the variables to be processed in an input file, FR be the memory required to store the entire contents of a single record of each of the variables to be processed in an input file, VR be the memory required to store a single record of the largest record variable to be processed in an input file, VT be the memory required to store the largest variable to be processed in an input file, VI be the memory required to store the largest variable which is not processed, but is copied from the initial file to the output file. All operators require MI = VI during the initial copying of variables from the first input file to the output file. This is the initial (and transient) memory demand. The sustained memory demand is that memory required by the operators during the processing (i.e., averaging, concatenation) phase which lasts until all the input files have been processed. The operators have the following memory requirements: ncrcat requires MS <= VR. ncecat requires MS <= VT. ncra requires MS = 2FR + VR. nces requires MS = 2FT + VT. ncbo requires MS <= 3VT (both input variables and the output variable). ncflint requires MS <= 3VT (both input variables and the output variable). ncpdq requires MS <= 2VT (one input variable and the output variable). ncwa requires MS <= 8VT (see below). Note that only variables that are processed, e.g., averaged, concatenated, or differenced, contribute to MS. Variables which do not appear in the output file (see Subsetting Files) are never read and contribute nothing to the memory requirements.

Further note that some operators perform internal type-promotion on some variables prior to arithmetic (see Type Conversion). For example, ncra and nces both promote integer types to double-precision floating-point prior to arithmetic, then perform the arithmetic, then demote back to the original integer type after arithmetic. This preserves the on-disk storage type while obtaining the accuracy advantages of floating-point arithmetic. Since version 4.3.6 (released in September, 2013), NCO also by default converts single-precision floating-point to double-precision prior to arithmetic, which incurs the same RAM penalty. Hence, the sustained memory required for integer variables and single-precision floats are two or four-times their on-disk, uncompressed, unpacked sizes if they meet the rules for automatic internal promotion. Put another way, disabling auto-promotion of single-precision variables (with ‘--flt’) considerably reduces the RAM footprint of arithmetic operators.

The ‘--open_ram’ switch (and switches that invoke it like ‘--ram_all’ and ‘--diskless_all’) incurs a RAM penalty. These switches cause each input file to be copied to RAM upon opening. Hence any operator invoking these switches utilizes an additional FT of RAM (i.e., MS += FT). See RAM disks for further details.

ncwa consumes between two and seven times the memory of a variable in order to process it. Peak consumption occurs when storing simultaneously in memory one input variable, one tally array, one input weight, one conformed/working weight, one weight tally, one input mask, one conformed/working mask, and one output variable. When invoked, the weighting and masking features contribute up to three-sevenths and two-sevenths of these requirements apiece. If weights and masks are not specified (i.e., no ‘-w’ or ‘-a’ options) then ncwa requirements drop to MS <= 3VT (one input variable, one tally array, and the output variable).

The above memory requirements must be multiplied by the number of threads thr_nbr (see OpenMP Threading). If this causes problems then reduce (with ‘-t thr_nbr’) the number of threads.

2.9.2 Memory for ncap2

ncap2 has unique memory requirements due its ability to process arbitrarily long scripts of any complexity. All scripts acceptable to ncap2 are ultimately processed as a sequence of binary or unary operations. ncap2 requires MS <= 2VT under most conditions. An exception to this is when left hand casting (see Left hand casting) is used to stretch the size of derived variables beyond the size of any input variables. Let VC be the memory required to store the largest variable defined by left hand casting. In this case, MS <= 2VC.

ncap2 scripts are complete dynamic and may be of arbitrary length. A script that contains many thousands of operations, may uncover a slow memory leak even though each single operation consumes little additional memory. Memory leaks are usually identifiable by their memory usage signature. Leaks cause peak memory usage to increase monotonically with time regardless of script complexity. Slow leaks are very difficult to find. Sometimes a malloc() (or new[]) failure is the only noticeable clue to their existence. If you have good reasons to believe that a memory allocation failure is ultimately due to an NCO memory leak (rather than inadequate RAM on your system), then we would be very interested in receiving a detailed bug report.

2.10 Performance

An overview of NCO capabilities as of about 2006 is in Zender, C. S. (2008), “Analysis of Self-describing Gridded Geoscience Data with netCDF Operators (NCO)”, Environ. Modell. Softw., doi:10.1016/j.envsoft.2008.03.004. This paper is also available at http://dust.ess.uci.edu/ppr/ppr_Zen08.pdf.

NCO performance and scaling for arithmetic operations is described in Zender, C. S., and H. J. Mangalam (2007), “Scaling Properties of Common Statistical Operators for Gridded Datasets”, Int. J. High Perform. Comput. Appl., 21(4), 485-498, doi:10.1177/1094342007083802. This paper is also available at http://dust.ess.uci.edu/ppr/ppr_ZeM07.pdf.

It is helpful to be aware of the aspects of NCO design that can limit its performance:

1. No data buffering is performed during nc_get_var and nc_put_var operations. Hyperslabs too large to hold in core memory will suffer substantial performance penalties because of this.
2. Since coordinate variables are assumed to be monotonic, the search for bracketing the user-specified limits should employ a quicker algorithm, like bisection, than the two-sided incremental search currently implemented.
3. C_format, FORTRAN_format, signedness, scale_format and add_offset attributes are ignored by ncks when printing variables to screen.
4. In the late 1990s it was discovered that some random access operations on large files on certain architectures (e.g., UNICOS) were much slower with NCO than with similar operations performed using languages that bypass the netCDF interface (e.g., Yorick). This may have been a penalty of unnecessary byte-swapping in the netCDF interface. It is unclear whether such problems exist in present day (2007) netCDF/NCO environments, where unnecessary byte-swapping has been reduced or eliminated.

3 Shared Features

Many features have been implemented in more than one operator and are described here for brevity. The description of each feature is preceded by a box listing the operators for which the feature is implemented. Command line switches for a given feature are consistent across all operators wherever possible. If no “key switches” are listed for a feature, then that particular feature is automatic and cannot be controlled by the user.

3.1 Internationalization

NCO support for internationalization of textual input and output (e.g., Warning messages) is nascent. We introduced the first foreign language string catalogues (French and Spanish) in 2004, yet did not activate these in distributions because the catalogues were nearly empty. We seek volunteers to populate our templates with translations for their favorite languages.

3.2 Metadata Optimization

NCO supports padding headers to improve the speed of future metadata operations. Use the ‘--hdr_pad’ and ‘--header_pad’ switches to request that hdr_pad bytes be inserted into the metadata section of the output file. Future metadata expansions will not incur the netCDF3 performance penalty of copying the entire output file unless the expansion exceeds the amount of header padding exceeded. This can be beneficial when it is known that some metadata will be added at a future date. The operators which benefit most from judicious use of header padding are ncatted and ncrename, since they only alter metadata.

This optimization exploits the netCDF library nc__enddef() function, which behaves differently with different versions of netCDF. It will improve speed of future metadata expansion with CLASSIC and 64bit netCDF files, though not necessarily with NETCDF4 files, i.e., those created by the netCDF interface to the HDF5 library (see File Formats and Conversion).

3.3 OpenMP Threading

NCO supports shared memory parallelism (SMP) when compiled with an OpenMP-enabled compiler. Threads requests and allocations occur in two stages. First, users may request a specific number of threads thr_nbr with the ‘-t’ switch (or its long option equivalents, ‘--thr_nbr’, ‘--threads’, and ‘--omp_num_threads’). If not user-specified, OpenMP obtains thr_nbr from the OMP_NUM_THREADS environment variable, if present, or from the OS, if not.

NCO may modify thr_nbr according to its own internal settings before it requests any threads from the system. Certain operators contain hard-code limits to the number of threads they request. We base these limits on our experience and common sense, and to reduce potentially wasteful system usage by inexperienced users. For example, ncrcat is extremely I/O-intensive so we restrict thr_nbr <= 2 for ncrcat. This is based on the notion that the best performance that can be expected from an operator which does no arithmetic is to have one thread reading and one thread writing simultaneously. In the future (perhaps with netCDF4), we hope to demonstrate significant threading improvements with operators like ncrcat by performing multiple simultaneous writes.

Compute-intensive operators (ncap2, ncwa and ncpdq) benefit most from threading. The greatest increases in throughput due to threading occur on large datasets where each thread performs millions, at least, of floating-point operations. Otherwise, the system overhead of setting up threads probably outweighs the speed enhancements due to SMP parallelism. However, we have not yet demonstrated that the SMP parallelism scales beyond four threads for these operators. Hence we restrict thr_nbr <= 4 for all operators. We encourage users to play with these limits (edit file nco_omp.c) and send us their feedback.

Once the initial thr_nbr has been modified for any operator-specific limits, NCO requests the system to allocate a team of thr_nbr threads for the body of the code. The operating system then decides how many threads to allocate based on this request. Users may keep track of this information by running the operator with dbg_lvl > 0.

By default, threaded operators attach one global attribute, nco_openmp_thread_number, to any file they create or modify. This attribute contains the number of threads the operator used to process the input files. This information helps to verify that the answers with threaded and non-threaded operators are equal to within machine precision. This information is also useful for benchmarking.

3.4 Command Line Options

NCO achieves flexibility by using command line options. These options are implemented in all traditional UNIX commands as single letter switches, e.g., ‘ls -l’. For many years NCO used only single letter option names. In late 2002, we implemented GNU/POSIX extended or long option names for all options. This was done in a backward compatible way such that the full functionality of NCO is still available through the familiar single letter options. Many features of NCO introduced since 2002 now require the use of long options, simply because we have nearly run out of single letter options. More importantly, mnemonics for single letter options are often non-intuitive so that long options provide a more natural way of expressing intent.

Extended options, also called long options, are implemented using the system-supplied getopt.h header file, if possible. This provides the getopt_long function to NCO ¹⁸.

The syntax of short options (single letter options) is -key value (dash-key-space-value). Here, key is the single letter option name, e.g., ‘-D 2’.

The syntax of long options (multi-letter options) is --long_name value (dash-dash-key-space-value), e.g., ‘--dbg_lvl 2’ or --long_name=value (dash-dash-key-equal-value), e.g., ‘--dbg_lvl=2’. Thus the following are all valid for the ‘-D’ (short version) or ‘--dbg_lvl’ (long version) command line option.

ncks -D 3 in.nc        # Short option, preferred form
ncks -D3 in.nc         # Short option, alternate form
ncks --dbg_lvl=3 in.nc # Long option, preferred form
ncks --dbg_lvl 3 in.nc # Long option, alternate form

The third example is preferred for two reasons. First, ‘--dbg_lvl’ is more specific and less ambiguous than ‘-D’. The long option format makes scripts more self documenting and less error-prone. Often long options are named after the source code variable whose value they carry. Second, the equals sign = joins the key (i.e., long_name) to the value in an uninterruptible text block. Experience shows that users are less likely to mis-parse commands when restricted to this form.

3.4.1 Truncating Long Options

GNU implements a superset of the POSIX standard. Their superset accepts any unambiguous truncation of a valid option:

ncks -D 3 in.nc        # Short option
ncks --dbg_lvl=3 in.nc # Long option, full form
ncks --dbg=3 in.nc     # Long option, OK unambiguous truncation
ncks --db=3 in.nc      # Long option, OK unambiguous truncation
ncks --d=3 in.nc       # Long option, ERROR ambiguous truncation

The first four examples are equivalent and will work as expected. The final example will exit with an error since ncks cannot disambiguate whether ‘--d’ is intended as a truncation of ‘--dbg_lvl’, of ‘--dimension’, or of some other long option.

NCO provides many long options for common switches. For example, the debugging level may be set in all operators with any of the switches ‘-D’, ‘--debug-level’, or ‘--dbg_lvl’. This flexibility allows users to choose their favorite mnemonic. For some, it will be ‘--debug’ (an unambiguous truncation of ‘--debug-level’, and other will prefer ‘--dbg’. Interactive users usually prefer the minimal amount of typing, i.e., ‘-D’. We recommend that re-usable scripts employ long options to facilitate self-documentation and maintainability.

This manual generally uses the short option syntax in examples. This is for historical reasons and to conserve space in printed output. Users are expected to pick the unambiguous truncation of each option name that most suits their taste.

3.4.2 Multi-arguments

As of NCO version 4.6.2 (November, 2016), NCO accepts multiple key-value pair options for a single feature to be joined together into a single extended argument called a multi-argument. Only four NCO features accept multiple key-value pairs that can be aggregated into multi-arguments. These features are: Global Attribute Addition options indicated via ‘--gaa’ (see Global Attribute Addition); Image Manipulation indicated via ‘--trr’¹⁹, Precision-Preserving Compression options indicated via ‘--ppc’ (see Precision-Preserving Compression); and Regridding options are indicated via ‘--rgr’ (see Regridding). Arguments to these four indicator options take the form of key-value pairs, e.g., ‘--rgr key=val’. These four features have so many options that making each key its own command line option would pollute the namespace of NCO’s global options. Yet supplying multiple options to each indicator option one-at-a-time can result in command lines overpopulated with indicator switches (e.g., ‘--rgr’):

ncks --rgr grd_ttl='Title' --rgr grid=grd.nc --rgr latlon=129,256 \
     --rgr lat_typ=fv --rgr lon_typ=grn_ctr ...

Multi-arguments combine all the indicator options into one option that receives a single argument that comprises all the original arguments glued together by a delimiter, which is, by default, ‘#’. Thus the multi-argument version of the above example is

ncks --rgr grd_ttl='Title'#grid=grd.nc#latlon=129,256#lat_typ=fv#lon_typ=grn_ctr

Note the aggregation of all key=val pairs into a single argument. NCO simply splits this argument at each delimiter, and processes the sub-arguments as if they had been passed with their own indicator option. Multi-arguments produce the same results, and may be mixed with, traditional indicator options supplied one-by-one.

As mentioned previously, the multi-argument delimiter string is, by default, the hash-sign ‘#’. When any key=val pair contains the default delimiter, the user must specify a custom delimiter string so that options are parsed correctly. The options to change the multi-argument delimiter string are ‘--mta_dlm=delim_string’ or ‘--dlm_mta=delim_string’, where delim_string can be any single or multi-character string that (1) is not contained in any key or val string; and (2) will not confuse the shell. For example, to use multi-arguments to pass a string that includes the hash symbol (the default delimiter is ‘#’), one must also change the delimiter so something besides hash, e.g., a colon ‘:’:

ncks --dlm=":" --gaa foo=bar:foo2=bar2:foo3,foo4="hash # is in value" 
ncks --dlm=":" --gaa foo=bar:foo2=bar2:foo3,foo4="Thu Sep 15 13\:03\:18 PDT 2016"
ncks --dlm="csz" --gaa foo=barcszfoo2=bar2cszfoo3,foo4="Long text"

In the second example, the colons that are escaped with the backslash become literal characters. Many characters have special shell meanings and so must be escaped by a single or double backslash or enclosed in single quotes to prevent interpolation. These special characters include ‘:’, ‘$’, ‘%’, ‘*’, ‘@’, and ‘&’. If val is a long text string that could contain the default delimiter, then delimit with a unique multi-character string such as ‘csz’ in the third example.

3.5 Specifying Input Files

It is important that users be able to specify multiple input files without typing every filename in full, often a tedious task even by graduate student standards. There are four different ways of specifying input files to NCO: explicitly typing each, using UNIX shell wildcards, and using the NCO ‘-n’ and ‘-p’ switches (or their long option equivalents, ‘--nintap’ or ‘--pth’ and ‘--path’, respectively). Techniques to augment these methods to specify arbitrary numbers (e.g., thousands) and patterns of filenames are discussed separately (see Large Numbers of Files).

To illustrate these methods, consider the simple problem of using ncra to average five input files, 85.nc, 86.nc, … 89.nc, and store the results in 8589.nc. Here are the four methods in order. They produce identical answers.

ncra 85.nc 86.nc 87.nc 88.nc 89.nc 8589.nc
ncra 8[56789].nc 8589.nc
ncra 8?.nc 8589.nc
ncra -p input-path 85.nc 86.nc 87.nc 88.nc 89.nc 8589.nc
ncra -n 5,2,1 85.nc 8589.nc

The first method (explicitly specifying all filenames) works by brute force. The second method relies on the operating system shell to glob (expand) the regular expression 8[56789].nc. The shell then passes the valid filenames (those which match the regular expansion) to ncra. In this case ncra never knows that a regular expression was used, because the shell intercepts and expands and matches the regular expression before ncra is actually invoked. The third method is uses globbing with a different regular expression that is less safe (it will also match unwanted files such as 81.nc and 8Z.nc if present). The fourth method uses the ‘-p input-path’ argument to specify the directory where all the input files reside. NCO prepends input-path (e.g., /data/username/model) to all input-files (though not to output-file). Thus, using ‘-p’, the path to any number of input files need only be specified once. Note input-path need not end with ‘/’; the ‘/’ is automatically generated if necessary.

The last method passes (with ‘-n’) syntax concisely describing the entire set of filenames ²⁰. This option is only available with the multi-file operators: ncra, ncrcat, nces, and ncecat. By definition, multi-file operators are able to process an arbitrary number of input-files. This option is very useful for abbreviating lists of filenames representable as alphanumeric_prefix+numeric_suffix+.+filetype where alphanumeric_prefix is a string of arbitrary length and composition, numeric_suffix is a fixed width field of digits, and filetype is a standard filetype indicator. For example, in the file ccm3_h0001.nc, we have alphanumeric_prefix = ccm3_h, numeric_suffix = 0001, and filetype = nc.

NCO decodes lists of such filenames encoded using the ‘-n’ syntax. The simpler (three-argument) ‘-n’ usage takes the form -n file_number,digit_number,numeric_increment where file_number is the number of files, digit_number is the fixed number of numeric digits comprising the numeric_suffix, and numeric_increment is the constant, integer-valued difference between the numeric_suffix of any two consecutive files. The value of alphanumeric_prefix is taken from the input file, which serves as a template for decoding the filenames. In the example above, the encoding -n 5,2,1 along with the input file name 85.nc tells NCO to construct five (5) filenames identical to the template 85.nc except that the final two (2) digits are a numeric suffix to be incremented by one (1) for each successive file. Currently filetype may be either be empty, nc, h5, cdf, hdf, hd5, or he5. If present, these filetype suffixes (and the preceding .) are ignored by NCO as it uses the ‘-n’ arguments to locate, evaluate, and compute the numeric_suffix component of filenames.

Recently the ‘-n’ option has been extended to allow convenient specification of filenames with “circular” characteristics. This means it is now possible for NCO to automatically generate filenames which increment regularly until a specified maximum value, and then wrap back to begin again at a specified minimum value. The corresponding ‘-n’ usage becomes more complex, taking one or two additional arguments for a total of four or five, respectively: -n file_number,digit_number,numeric_increment[,numeric_max[,numeric_min]] where numeric_max, if present, is the maximum integer-value of numeric_suffix and numeric_min, if present, is the minimum integer-value of numeric_suffix. Consider, for example, the problem of specifying non-consecutive input files where the filename suffixes end with the month index. In climate modeling it is common to create summertime and wintertime averages which contain the averages of the months June–July–August, and December–January–February, respectively:

ncra -n 3,2,1 85_06.nc 85_0608.nc
ncra -n 3,2,1,12 85_12.nc 85_1202.nc
ncra -n 3,2,1,12,1 85_12.nc 85_1202.nc

The first example shows that three arguments to the ‘-n’ option suffice to specify consecutive months (06, 07, 08) which do not “wrap” back to a minimum value. The second example shows how to use the optional fourth and fifth elements of the ‘-n’ option to specify a wrap value. The fourth argument to ‘-n’, when present, specifies the maximum integer value of numeric_suffix. In the example the maximum value is 12, and will be formatted as 12 in the filename string. The fifth argument to ‘-n’, when present, specifies the minimum integer value of numeric_suffix. The default minimum filename suffix is 1, which is formatted as 01 in this case. Thus the second and third examples have the same effect, that is, they automatically generate, in order, the filenames 85_12.nc, 85_01.nc, and 85_02.nc as input to NCO.

As of NCO version 4.5.2 (September, 2015), NCO supports an optional sixth argument to ‘-n’, the month-indicator. The month-indicator affirms to NCO that the right-most digits being manipulated in the generated filenames correspond to month numbers (with January formatted as 01 and December as 12). Moreover, it assumes digits to the left of the month are the year. The full (six-argument) ‘-n’ usage takes the form -n file_number,digit_number,month_increment,max_month,min_month,‘yyyymm’. The ‘yyyymm’ string is a clunky way (can you think of a clearer way?) to tell NCO to enumerate files in year-month mode. When present, ‘yyyymm’ string causes NCO to automatically generate series of filenames whose right-most two digits increment from min_month by month_increment up to max_month and then the leftmost digits (i.e., the year) increment by one, and the whole process is reapeated until the file_number filenames are generated.

ncrcat -n 3,6,1,12,1         198512.nc 198512_198502.nc
ncrcat -n 3,6,1,12,1,yyyymm  198512.nc 198512_198602.nc
ncrcat -n 3,6,1,12,12,yyyymm 198512.nc 198512_198712.nc

The first command concatenates three files (198512.nc, 198501.nc, 198502.nc) into the output file. The second command concatenates three files (198512.nc, 198601.nc, 198602.nc). The ‘yyyymm’-indicator causes the left-most digits to increment each time the right-most two digits reach their maximum and then wrap. The first command does not have the indicator so it is always 1985. The third command concatenates three files (198512.nc, 198612.nc, 198712.nc).

3.6 Specifying Output Files

NCO commands produce no more than one output file, fl_out. Traditionally, users specify fl_out as the final argument to the operator, following all input file names. This is the positional argument method of specifying input and ouput file names. The positional argument method works well in most applications. NCO also supports specifying fl_out using the command line switch argument method, ‘-o fl_out’.

Specifying fl_out with a switch, rather than as a positional argument, allows fl_out to precede input files in the argument list. This is particularly useful with multi-file operators for three reasons. Multi-file operators may be invoked with hundreds (or more) filenames. Visual or automatic location of fl_out in such a list is difficult when the only syntactic distinction between input and output files is their position. Second, specification of a long list of input files may be difficult (see Large Numbers of Files). Making the input file list the final argument to an operator facilitates using xargs for this purpose. Some alternatives to xargs are heinous and undesirable. Finally, many users are more comfortable specifying output files with ‘-o fl_out’ near the beginning of an argument list. Compilers and linkers are usually invoked this way.

Users should specify fl_out using either (not both) method. If fl_out is specified twice (once with the switch and once as the last positional argument), then the positional argument takes precedence.

3.7 Accessing Remote Files

All NCO operators can retrieve files from remote sites as well as from the local file system. A remote site can be an anonymous FTP server, a machine on which the user has rcp, scp, or sftp privileges, NCAR’s Mass Storage System (MSS), or an OPeNDAP server. Examples of each are given below, following a brief description of the particular access protocol.

To access a file via an anonymous FTP server, simply supply the remote file’s URL. Anonymous FTP usually requires no further credentials, e.g., no .netrc file is necessary. FTP is an intrinsically insecure protocol because it transfers passwords in plain text format. Users should access sites using anonymous FTP, or better yet, secure FTP (SFTP, see below) when possible. Some FTP servers require a login/password combination for a valid user account. NCO allows transactions that require additional credentials so long as the required information is stored in the .netrc file. Usually this information is the remote machine name, login, and password, in plain text, separated by those very keywords, e.g.,

machine dust.ess.uci.edu login zender password bushlied

Eschew using valuable passwords for FTP transactions, since .netrc passwords are potentially exposed to eavesdropping software ²¹.

SFTP, i.e., secure FTP, uses SSH-based security protocols that solve the security issues associated with plain FTP. NCO supports SFTP protocol access to files specified with a homebrew syntax of the form

sftp://machine.domain.tld:/path/to/filename

Note the second colon following the top-level-domain, tld. This syntax is a hybrid between an FTP URL and standard remote file syntax.

To access a file using rcp or scp, specify the Internet address of the remote file. Of course in this case you must have rcp or scp privileges which allow transparent (no password entry required) access to the remote machine. This means that ~/.rhosts or ~/ssh/authorized_keys must be set accordingly on both local and remote machines.

To access a file on a High Performance Storage System (HPSS) (such as that at NCAR, ECMWF, LANL, DKRZ, LLNL) specify the full HPSS pathname of the remote file. NCO will attempt to detect whether the local machine has direct (synchronous) HPSS access. In this case, NCO attempts to use the Hierarchical Storage Interface (HSI) command hsi get ²².

The following examples show how one might analyze files stored on remote systems.

ncks -l . ftp://dust.ess.uci.edu/pub/zender/nco/in.nc
ncks -l . sftp://dust.ess.uci.edu:/home/ftp/pub/zender/nco/in.nc
ncks -l . dust.ess.uci.edu:/home/zender/nco/data/in.nc
ncks -l . /ZENDER/nco/in.nc
ncks -l . /home/zender/nco/in.nc
ncks -l . http://thredds-test.ucar.edu/thredds/dodsC/testdods/in.nc 

The first example works verbatim if your system is connected to the Internet and is not behind a firewall. The second example works if you have sftp access to the machine dust.ess.uci.edu. The third example works if you have rcp or scp access to the machine dust.ess.uci.edu. The fourth and fifth examples work on NCAR computers with local access to the HPSS hsi get command ²³. The sixth command works if your local version of NCO is OPeNDAP-enabled (this is fully described in OPeNDAP), or if the remote file is accessible via wget. The above commands can be rewritten using the ‘-p input-path’ option as follows:

ncks -p ftp://dust.ess.uci.edu/pub/zender/nco -l . in.nc
ncks -p sftp://dust.ess.uci.edu:/home/ftp/pub/zender/nco -l . in.nc
ncks -p dust.ess.uci.edu:/home/zender/nco -l . in.nc
ncks -p /ZENDER/nco -l . in.nc
ncks -p /home/zender/nco -l . in.nc # HPSS
ncks -p http://thredds-test.ucar.edu/thredds/dodsC/testdods \ 
     -l . in.nc

Using ‘-p’ is recommended because it clearly separates the input-path from the filename itself, sometimes called the stub. When input-path is not explicitly specified using ‘-p’, NCO internally generates an input-path from the first input filename. The automatically generated input-path is constructed by stripping the input filename of everything following the final ‘/’ character (i.e., removing the stub). The ‘-l output-path’ option tells NCO where to store the remotely retrieved file. It has no effect on locally-retrieved files, or on the output file. Often the path to a remotely retrieved file is quite different than the path on the local machine where you would like to store the file. If ‘-l’ is not specified then NCO internally generates an output-path by simply setting output-path equal to input-path stripped of any machine names. If ‘-l’ is not specified and the remote file resides on the NCAR HPSS system, then the leading character of input-path, ‘/’, is also stripped from output-path. Specifying output-path as ‘-l ./’ tells NCO to store the remotely retrieved file and the output file in the current directory. Note that ‘-l .’ is equivalent to ‘-l ./’ though the latter is syntactically more clear.

3.7.1 OPeNDAP

The Distributed Oceanographic Data System (DODS) provides useful replacements for common data interface libraries like netCDF. The DODS versions of these libraries implement network transparent access to data via a client-server data access protocol that uses the HTTP protocol for communication. Although DODS-technology originated with oceanography data, it applyies to virtually all scientific data. In recognition of this, the data access protocol underlying DODS (which is what NCO cares about) has been renamed the Open-source Project for a Network Data Access Protocol, OPeNDAP. We use the terms DODS and OPeNDAP interchangeably, and often write OPeNDAP/DODS for now. In the future we will deprecate DODS in favor of DAP or OPeNDAP, as appropriate ²⁴.

NCO may be DAP-enabled by linking NCO to the OPeNDAP libraries. This is described in the OPeNDAP documentation and automagically implemented in NCO build mechanisms ²⁵. The ./configure mechanism automatically enables NCO as OPeNDAP clients if it can find the required OPeNDAP libraries ²⁶. in the usual locations. The $DODS_ROOT environment variable may be used to override the default OPeNDAP library location at NCO compile-time. Building NCO with bld/Makefile and the command make DODS=Y adds the (non-intuitive) commands to link to the OPeNDAP libraries installed in the $DODS_ROOT directory. The file doc/opendap.sh contains a generic script intended to help users install OPeNDAP before building NCO. The documentation at the OPeNDAP Homepage is voluminous. Check there and on the DODS mail lists. to learn more about the extensive capabilities of OPeNDAP ²⁷.

Once NCO is DAP-enabled the operators are OPeNDAP clients. All OPeNDAP clients have network transparent access to any files controlled by a OPeNDAP server. Simply specify the input file path(s) in URL notation and all NCO operations may be performed on remote files made accessible by a OPeNDAP server. This command tests the basic functionality of OPeNDAP-enabled NCO clients:

% ncks -O -o ~/foo.nc -C -H -v one -l /tmp \
  -p http://thredds-test.ucar.edu/thredds/dodsC/testdods in.nc
% ncks -H -v one ~/foo.nc
one = 1

The one = 1 outputs confirm (first) that ncks correctly retrieved data via the OPeNDAP protocol and (second) that ncks created a valid local copy of the subsetted remote file. With minor changes to the above command, netCDF4 can be used as both the input and output file format:

% ncks -4 -O -o ~/foo.nc -C -H -v one -l /tmp \
  -p http://thredds-test.ucar.edu/thredds/dodsC/testdods in_4.nc
% ncks -H -v one ~/foo.nc
one = 1

And, of course, OPeNDAP-enabled NCO clients continue to support orthogonal features such as UDUnits (see UDUnits Support):

% ncks -u -C -H -v wvl -d wvl,'0.4 micron','0.7 micron' \
  -p http://thredds-test.ucar.edu/thredds/dodsC/testdods in_4.nc
% wvl[0]=5e-07 meter

The next command is a more advanced example which demonstrates the real power of OPeNDAP-enabled NCO clients. The ncwa client requests an equatorial hyperslab from remotely stored NCEP reanalyses data of the year 1969. The NOAA OPeNDAP server (hopefully!) serves these data. The local ncwa client then computes and stores (locally) the regional mean surface pressure (in Pa).

ncwa -C -a lat,lon,time -d lon,-10.,10. -d lat,-10.,10. -l /tmp -p \
http://www.esrl.noaa.gov/psd/thredds/dodsC/Datasets/ncep.reanalysis.dailyavgs/surface \
  pres.sfc.1969.nc ~/foo.nc

All with one command! The data in this particular input file also happen to be packed (see Methods and functions), although this complication is transparent to the user since NCO automatically unpacks data before attempting arithmetic.

NCO obtains remote files from the OPeNDAP server (e.g., www.cdc.noaa.gov) rather than the local machine. Input files are first copied to the local machine, then processed. The OPeNDAP server performs data access, hyperslabbing, and transfer to the local machine. This allows the I/O to appear to NCO as if the input files were local. The local machine performs all arithmetic operations. Only the hyperslabbed output data are transferred over the network (to the local machine) for the number-crunching to begin. The advantages of this are obvious if you are examining small parts of large files stored at remote locations.

Natually there are many versions of OPeNDAP servers supplying data and bugs in the server can appear to be bugs in NCO. However, with very few exceptions ²⁸ an NCO command that works on a local file must work across an OPeNDAP connection or else there is a bug in the server. This is because NCO does nothing special to handle files served by OPeNDAP, the whole process is (supposed to be) completely transparent to the client NCO software. Therefore it is often useful to try NCO commands on various OPeNDAP servers in order to isolate whether a problem may be due to a bug in the OPeNDAP server on a particular machine. For this purpose, one might try variations of the following commands that access files on public OPeNDAP servers:

# Strided access to HDF5 file
ncks -v Time -d Time,0,10,2 http://eosdap.hdfgroup.uiuc.edu:8080/opendap/data/NASAFILES/hdf5/BUV-Nimbus04_L3zm_v01-00-2012m0203t144121.h5
# Strided access to netCDF3 file
ncks -O -D 1 -d time,1 -d lev,0 -d lat,0,100,10 -d lon,0,100,10 -v u_velocity http://nomads.ncep.noaa.gov:9090/dods/rtofs/rtofs_global20140303/rtofs_glo_2ds_forecast_daily_prog ~/foo.nc

These servers were operational at the time of writing, March 2014. Unfortunately, administrators often move or rename path directories. Recommendations for additional public OPeNDAP servers on which to test NCO are welcome.

3.8 Retaining Retrieved Files

In order to conserve local file system space, files retrieved from remote locations are automatically deleted from the local file system once they have been processed. Many NCO operators were constructed to work with numerous large (e.g., 200 MB) files. Retrieval of multiple files from remote locations is done serially. Each file is retrieved, processed, then deleted before the cycle repeats. In cases where it is useful to keep the remotely-retrieved files on the local file system after processing, the automatic removal feature may be disabled by specifying ‘-R’ on the command line.

Invoking -R disables the default printing behavior of ncks. This allows ncks to retrieve remote files without automatically trying to print them. See ncks netCDF Kitchen Sink, for more details.

Note that the remote retrieval features of NCO can always be used to retrieve any file, including non-netCDF files, via SSH, anonymous FTP, or msrcp. Often this method is quicker than using a browser, or running an FTP session from a shell window yourself. For example, say you want to obtain a JPEG file from a weather server.

ncks -R -p ftp://weather.edu/pub/pix/jpeg -l . storm.jpg

In this example, ncks automatically performs an anonymous FTP login to the remote machine and retrieves the specified file. When ncks attempts to read the local copy of storm.jpg as a netCDF file, it fails and exits, leaving storm.jpg in the current directory.

If your NCO is DAP-enabled (see OPeNDAP), then you may use NCO to retrieve any files (including netCDF, HDF, etc.) served by an OPeNDAP server to your local machine. For example,

ncks -R -l . -p \
http://www.esrl.noaa.gov/psd/thredds/dodsC/Datasets/ncep.reanalysis.dailyavgs/surface \
  pres.sfc.1969.nc

It may occasionally be useful to use NCO to transfer files when your other preferred methods are not available locally.

3.9 File Formats and Conversion

All NCO operators support (read and write) all three (or four, depending on how one counts) file formats supported by netCDF4. The default output file format for all operators is the input file format. The operators listed under “Availability” above allow the user to specify the output file format independent of the input file format. These operators allow the user to convert between the various file formats. (The operators ncatted and ncrename do not support these switches so they always write the output netCDF file in the same format as the input netCDF file.)

3.9.1 File Formats

netCDF supports five types of files: CLASSIC, 64BIT_OFFSET, 64BIT_DATA, NETCDF4, and NETCDF4_CLASSIC. The CLASSIC format is the traditional 32-bit offset written by netCDF2 and netCDF3. As of 2005, nearly all netCDF datasets were in CLASSIC format. The 64BIT_OFFSET (originally called plain old 64BIT) format was added in Fall, 2004. As of 2010, many netCDF datasets were in 64BIT_OFFSET format. As of 2013, an increasing number of netCDF datasets were in NETCDF4_CLASSIC format. The 64BIT_DATA (aka CDF5 or PNETCDF) format was added in January, 2016.

The NETCDF4 format uses HDF5 as the file storage layer. The files are (usually) created, accessed, and manipulated using the traditional netCDF3 API (with numerous extensions). The NETCDF4_CLASSIC format refers to netCDF4 files created with the NC_CLASSIC_MODEL mask. Such files use HDF5 as the back-end storage format (unlike netCDF3), though they incorporate only netCDF3 features. Hence NETCDF4_CLASSIC files are entirely readable by applications that use only the netCDF3 API (though the applications must be linked with the netCDF4 library). NCO must be built with netCDF4 to write files in the new NETCDF4 and NETCDF4_CLASSIC formats, and to read files in these formats. Datasets in the default CLASSIC or the newer 64BIT_OFFSET formats have maximum backwards-compatibility with older applications. NCO has deep support for NETCDF4 formats. If performance or disk-space as important as backwards compatibility, then use NETCDF4_CLASSIC instead of CLASSIC format files. As of 2014, NCO support for the NETCDF4 format is nearly complete and the most powerful and disk/RAM efficient workflows will utilize this format.

As mentioned above, all operators write use the input file format for output files unless told otherwise. Toggling the short option ‘-6’ or the long option ‘--6’ or ‘--64bit_offset’ (or their key-value equivalent ‘--fl_fmt=64bit_offset’) produces the netCDF3 64-bit offset format named 64BIT_OFFSET. NCO must be built with netCDF 3.6 or higher to produce a 64BIT_OFFSET file. Using the ‘-4’ switch (or its long option equivalents ‘--4’ or ‘--netcdf4’), or setting its key-value equivalent ‘--fl_fmt=netcdf4’ produces a NETCDF4 file (i.e., with all supported HDF5 features). Using the ‘-7’ switch (or its long option equivalent ‘--7’ ²⁹, or setting its key-value equivalent ‘--fl_fmt=netcdf4_classic’ produces a NETCDF4_CLASSIC file (i.e., with all supported HDF5 features like compression and chunking but without groups or new atomic types). Operators given the ‘-3’ (or ‘--3’) switch without arguments will (attempt to) produce netCDF3 CLASSIC output, even from netCDF4 input files.

Note that NETCDF4 and NETCDF4_CLASSIC are the same binary format. The latter simply causes a writing application to fail if it attempts to write a NETCDF4 file that cannot be completely read by the netCDF3 library. Conversely, NETCDF4_CLASSIC indicates to a reading application that all of the file contents are readable with the netCDF3 library. NCO has supported reading/writing basic NETCDF4 and NETCDF4_CLASSIC files since October, 2005.

3.9.2 Determining File Format

Input files often end with the generic .nc suffix that leaves (perhaps by intention) the internal file format ambiguous. There are at least three ways to discover the internal format of a netCDF-supported file. These methods determine whether it is a classic (32-bit offset) or newer 64-bit offset netCDF3 format, or is a netCDF4 format. Each method returns the information using slightly different terminology that becomes easier to understand with practice.

First, examine the first line of global metadata output by ‘ncks -M’:

% ncks -M foo_3.nc
Summary of foo_3.nc: filetype = NC_FORMAT_CLASSIC, 0 groups ...
% ncks -M foo_6.nc
Summary of foo_6.nc: filetype = NC_FORMAT_64BIT_OFFSET, 0 groups ...
% ncks -M foo_5.nc
Summary of foo_5.nc: filetype = NC_FORMAT_CDF5, 0 groups ...
% ncks -M foo_7.nc
Summary of foo_7.nc: filetype = NC_FORMAT_NETCDF4_CLASSIC, 0 groups ...
% ncks -M foo_4.nc
Summary of foo_4.nc: filetype = NC_FORMAT_NETCDF4, 0 groups ...

This method requires a netCDF4-enabled NCO version 3.9.0+ (i.e., from 2007 or later). As of NCO version 4.4.0 (January, 2014), ncks will also print the extended or underlying format of the input file. The extended filetype will be one of the six underlying formats that are accessible through the netCDF API. These formats are NC_FORMATX_NC3 (classic and 64-bit versions of netCDF3 formats), NC_FORMATX_NC_HDF5 (classic and extended versions of netCDF4, and “pure” HDF5 format), NC_FORMATX_NC_HDF4 (HDF4 format), NC_FORMATX_PNETCDF (PnetCDF format), NC_FORMATX_DAP2 (accessed via DAP2 protocol), and NC_FORMATX_DAP4 (accessed via DAP4 protocol). For example,

% ncks -D 2 -M hdf.hdf
Summary of hdf.hdf: filetype = NC_FORMAT_NETCDF4 (representation of \
  extended/underlying filetype NC_FORMAT_HDF4), 0 groups ...
% ncks -D 2 -M http://thredds-test.ucar.edu/thredds/dodsC/testdods/in.nc
Summary of http://thredds-test.ucar.edu/thredds/dodsC/testdods/in.nc: \
  filetype = NC_FORMAT_CLASSIC (representation of extended/underlying \
  filetype NC_FORMATX_DAP2), 0 groups  
% ncks -D 2 -M foo_4.nc
Summary of foo_4.nc: filetype = NC_FORMAT_NETCDF4 (representation of \
  extended/underlying filetype NC_FORMAT_HDF5), 0 groups  

The extended filetype determines some of the capabilities that netCDF has to alter the file.

Second, query the file with ‘ncdump -k’:

% ncdump -k foo_3.nc
classic
% ncdump -k foo_6.nc
64-bit offset
% ncdump -k foo_5.nc
cdf5
% ncdump -k foo_7.nc
netCDF-4 classic model
% ncdump -k foo_4.nc
netCDF-4

This method requires a netCDF4-enabled netCDF 3.6.2+ (i.e., from 2007 or later).

The third option uses the POSIX-standard od (octal dump) command:

% od -An -c -N4 foo_3.nc
   C   D   F 001
% od -An -c -N4 foo_6.nc
   C   D   F 002
% od -An -c -N4 foo_5.nc
   C   D   F 005
% od -An -c -N4 foo_7.nc
 211   H   D   F
% od -An -c -N4 foo_4.nc
 211   H   D   F

This option works without NCO and ncdump. Values of ‘C D F 001’ and ‘C D F 002’ indicate 32-bit (classic) and 64-bit netCDF3 formats, respectively, while values of ‘211 H D F’ indicate either of the newer netCDF4 file formats.

3.9.3 File Conversion

Let us demonstrate converting a file from any netCDF-supported input format into any netCDF output format (subject to limits of the output format). Here the input file in.nc may be in any of these formats: netCDF3 (classic, 64bit_offset, 64bit_data), netCDF4 (classic and extended), HDF4, HDF5, HDF-EOS (version 2 or 5), and DAP. The switch determines the output format written in the comment: ³⁰

ncks --fl_fmt=classic in.nc foo_3.nc # netCDF3 classic
ncks --fl_fmt=64bit_offset in.nc foo_6.nc # netCDF3 64bit-offset
ncks --fl_fmt=64bit_data in.nc foo_5.nc # netCDF3 64bit-data
ncks --fl_fmt=cdf5 in.nc foo_5.nc # netCDF3 64bit-data
ncks --fl_fmt=netcdf4_classic in.nc foo_7.nc # netCDF4 classic
ncks --fl_fmt=netcdf4 in.nc foo_4.nc # netCDF4 
ncks -3 in.nc foo_3.nc # netCDF3 classic
ncks --3 in.nc foo_3.nc # netCDF3 classic
ncks -6 in.nc foo_6.nc # netCDF3 64bit-offset
ncks --64 in.nc foo_6.nc # netCDF3 64bit-offset
ncks -5 in.nc foo_5.nc # netCDF3 64bit-data
ncks --5 in.nc foo_5.nc # netCDF3 64bit-data
ncks -4 in.nc foo_4.nc # netCDF4 
ncks --4 in.nc foo_4.nc # netCDF4 
ncks -7 in.nc foo_7.nc # netCDF4 classic
ncks --7 in.nc foo_7.nc # netCDF4 classic

Of course since most operators support these switches, the “conversions” can be done at the output stage of arithmetic or metadata processing rather than requiring a separate step. Producing (netCDF3) CLASSIC or 64BIT_OFFSET or 64BIT_DATA files from NETCDF4_CLASSIC files always works.

3.9.4 Autoconversion

Because of the dearth of support for netCDF4 amongst tools and user communities (including the CF conventions), it is often useful to convert netCDF4 to netCDF3 for certain applications. Until NCO version 4.4.0 (January, 2014), producing netCDF3 files from netCDF4 files only worked if the input files contained no netCDF4-specific features (e.g., atomic types, multiple record dimensions, or groups). As of NCO version 4.4.0, ncks supports autoconversion of many netCDF4 features to their closest netCDF3-compatible representations. Since converting netCDF4 to netCDF3 results in loss of features, “automatic down-conversion” may be a more precise description of what we term autoconversion.

NCO employs three algorithms to downconvert netCDF4 to netCDF3:

1. Autoconversion of atomic types: Autoconversion automatically promotes NC_UBYTE to NC_SHORT, and NC_USHORT to NC_INT. It automatically demotes the three types NC_UINT, NC_UINT64, and NC_INT64 to NC_INT. And it converts NC_STRING to NC_CHAR. All numeric conversions work for attributes and variables of any rank. Two numeric types (NC_UBYTE and NC_USHORT) are promoted to types with greater range (and greater storage). This extra range is often not used so promotion perhaps conveys the wrong impression. However, promotion never truncates values or loses data (this perhaps justifies the extra storage). Three numeric types (NC_UINT, NC_UINT64 and NC_INT64) are demoted. Since the input range is larger than the output range, demotion can result in numeric truncation and thus loss of data. In such cases, it would possible to convert the data to floating-point values instead. If this feature interests you, please be the squeaky wheel and let us know.

   String conversions (to NC_CHAR) work for all attributes, but not for variables. This is because attributes are at most one-dimensional and may be of any size whereas variables require gridded dimensions that usually do not fit the ragged sizes of text strings. Hence scalar NC_STRING attributes are correctly converted to and stored as NC_CHAR attributes in the netCDF3 output file, but NC_STRING variables are not correctly converted. If this limitation annoys or enrages you, please let us know by being the squeaky wheel.

2. Convert multiple record dimensions to fixed-size dimensions. Many netCDF4 and HDF5 datasets have multiple unlimited dimensions. Since a netCDF3 file may have at most one unlimited dimension, all but possibly one unlimited dimension from the input file must be converted to fixed-length dimensions prior to storing netCDF4 input as netCDF3 output. By invoking --fix_rec_dmn all the user ensures the output file will adhere to netCDF3 conventions and the user need not know the names of the specific record dimensions to fix. See ncks netCDF Kitchen Sink for a description of the ‘--fix_rec_dmn’ option.
3. Flattening (removal) of groups. Many netCDF4 and HDF5 datasets have group hierarchies. Since a netCDF3 file may not have any groups, groups in the input file must be removed. This is also called “flattening” the hierarchical file. See Group Path Editing for a description of the GPE option ‘-G :’ to flatten files.

Putting the three algorithms together, one sees that the recipe to convert netCDF4 to netCDF4 becomes increasingly complex as the netCDF4 features in the input file become more elaborate:

# Convert file with netCDF4 atomic types
ncks -3 in.nc4 out.nc3
# Convert file with multiple record dimensions + netCDF4 atomic types
ncks -3 --fix_rec_dmn=all in.nc4 out.nc3
# Convert file with groups, multiple record dimensions + netCDF4 atomic types
ncks -3 -G : --fix_rec_dmn=all in.nc4 out.nc3

Future versions of NCO may automatically invoke the record dimension fixation and group flattening when converting to netCDF3 (rather than requiring it be specified manually). If this feature would interest you, please let us know.

3.10 Large File Support

NCO has Large File Support (LFS), meaning that NCO can write files larger than 2 GB on some 32-bit operating systems with netCDF libraries earlier than version 3.6. If desired, LFS support must be configured when both netCDF and NCO are installed. netCDF versions 3.6 and higher support 64-bit file addresses as part of the netCDF standard. We recommend that users ignore LFS support which is difficult to configure and is implemented in NCO only to support netCDF versions prior to 3.6. This obviates the need for configuring explicit LFS support in applications (such as NCO) that now support 64-bit files directly through the netCDF interface. See File Formats and Conversion for instructions on accessing the different file formats, including 64-bit files, supported by the modern netCDF interface.

If you are still interested in explicit LFS support for netCDF versions prior to 3.6, know that LFS support depends on a complex, interlocking set of operating system ³¹ and netCDF support issues. The netCDF LFS FAQ describes the various file size limitations imposed by different versions of the netCDF standard. NCO and netCDF automatically attempt to configure LFS at build time.

3.11 Subsetting Files

Subsetting variables refers to explicitly specifying variables and groups to be included or excluded from operator actions. Subsetting is controlled by the ‘-v var[,…]’ and ‘-x’ options for directly specifying variables. Specifying groups, whether in addition to or instead of variables, is quite similar and is controlled by the ‘-g grp[,…]’ and ‘-x’ options. A list of variables or groups to extract is specified following the ‘-v’ and ‘-g’ options, e.g., ‘-v time,lat,lon’ or ‘-g grp1,grp2’. Both options may be specified simultaneously and NCO will extract the intersection of the lists, i.e., only variables of the specified names found in groups of the specified names. The ‘--unn’ option causes NCO to extract the union, rather than the intersection, of the specified groups and variables. Not using the ‘-v’ or ‘-g’ option is equivalent to specifying all variables or groupp, respectively. The ‘-x’ option causes the list of variables specified with ‘-v’ to be excluded rather than extracted. Thus ‘-x’ saves typing when you only want to extract fewer than half of the variables in a file.

Variables or groups explicitly specified for extraction with ‘-v var[,…]’ or ‘-g grp[,…]’ must be present in the input file or an error will result. Variables explicitly specified for exclusion with ‘-x -v var[,…]’ need not be present in the input file. To accord with the sophistication of the underlying hierarchy, group subsetting is controlled by a few powerful yet subtle syntactical distinctions. When learning this syntax it is helpful to keep in mind the similarity between group hierarchies and directory structures.

As of NCO 4.4.4 (June, 2014), ncks (alone) supports an option to include specified groups yet exclude specified variables. The ‘--grp_xtr_var_xcl’ switch (with long option equivalent ‘--gxvx’) extracts all contents of groups given as arguments to ‘-g grp[,…]’, except for variables given as arguments to ‘-v var[,…]’. Use this when one or a few variables in hierarchical files are not to be extracted, and all other variables are. This is useful when coercing netCDF4 files into netCDF3 files such as with converting, flattening, or dismembering files (see Flattening Groups).

ncks --grp_xtr_var_xcl -g g1 -v v1 # Extract all of group g1 except v1

Two properties of subsetting, recursion and anchoring, are best illustrated by reminding the user of their UNIX equivalents. The UNIX command mv src dst moves src and all its subdirectories (and all their subdirectories etc.) to dst. In other words mv is, by default, recursive. In contrast, the UNIX command cp src dst moves src, and only src, to dst, If src is a directory, not a file, then that command fails. One must explicitly request to copy directories recursively, i.e., with cp -r src dst. In NCO recursive extraction (and copying) of groups is the default (like with mv, not with cp). Recursion is turned off by appending a trailing slash to the path.

These UNIX commands also illustrate a property we call anchoring. The command mv src dst moves (recursively) the source directory src to the destination directory dst. If src begins with the slash character then the specified path is relative to the root directory, otherwise the path is relative to the current working directory. In other words, an initial slash character anchors the subsequent path to the root directory. In NCO an initial slash anchors the path at the root group. Paths that begin and end with slash characters (e.g., //, /g1/, and /g1/g2/) are both anchored and non-recursive.

Consider the following commands, all of which may be assumed to end with ‘in.nc out.nc’:

ncks -g  g1  # Extract, recursively, all groups with a g1 component
ncks -g  g1/ # Extract, non-recursively, all groups terminating in g1
ncks -g /g1  # Extract, recursively, root group g1
ncks -g /g1/ # Extract, non-recursively root group g1
ncks -g //   # Extract, non-recursively the root group

The first command is probably the most useful and common. It would extract these groups, if present, and all their direct ancestors and children: /g1, /g2/g1, and /g3/g1/g2. In other words, the simplest form of ‘-g grp’ grabs all groups that (and their direct ancestors and children, recursively) that have grp as a complete component of their path. A simple string match is insufficient, grp must be a complete component (i.e., group name) in the path. The option ‘-g g1’ would not extract these groups because g1 is not a complete component of the path: /g12, /fg1, and /g1g1. The second command above shows how a terminating slash character / cancels the recursive copying of groups. An argument to ‘-g’ which terminates with a slash character extracts the group and its direct ancestors, but none of its children. The third command above shows how an initial slash character / anchors the argument to the root group. The third command would not extract the group /g2/g1 because the g1 group is not at the root level, but it would extract, any group /g1 at the root level and all its children, recursively. The fourth command is the non-recursive version of the third command. The fifth command is a special case of the fourth command.

As mentioned above, both ‘-v’ and ‘-g’ options may be specified simultaneously and NCO will, by default, extract the intersection of the lists, i.e., the specified variables found in the specified groups ³². The ‘--unn’ option causes NCO to extract the union, rather than the intersection, of the specified groups and variables. Consider the following commands (which may be assumed to end with ‘in.nc out.nc’):

# Intersection-mode subsetting (default)
ncks -g  g1  -v v1 # Yes: /g1/v1, /g2/g1/v1. No: /v1, /g2/v1
ncks -g /g1  -v v1 # Yes: /g1/v1, /g1/g2/v1. No: /v1, /g2/v1, /g2/g1/v1
ncks -g  g1/ -v v1 # Yes: /g1/v1, /g2/g1/v1. No: /v1, /g2/v1, /g1/g2/v1
ncks -v  g1/v1     # Yes: /g1/v1, /g2/g1/v1. No: /v1, /g2/v1, /g1/g2/v1
ncks -g /g1/ -v v1 # Yes: /g1/v1. No: /g2/g1/v1, /v1, /g2/v1 ...
ncks -v /g1/v1     # Yes: /g1/v1. No: /g2/g1/v1, /v1, /g2/v1 ...

# Union-mode subsetting (invoke with --unn or --union)
ncks -g  g1  -v v1 --unn # All variables in  g1 or progeny, or named v1
ncks -g /g1  -v v1 --unn # All variables in /g1 or progeny, or named v1
ncks -g  g1/ -v v1 --unn # All variables in  g1 or named v1
ncks -g /g1/ -v v1 --unn # All variables in /g1 or named v1

The first command (‘-g g1 -v v1’) extracts the variable v1 from any group named g1 or descendent g1. The second command extracts v1 from any root group named g1 and any descendent groups as well. The third and fourth commands are equivalent ways of extracting v1 only from the root group named g1 (not its descendents). The fifth and sixth commands are equivalent ways of extracting the variable v1 only from the root group named g1. Subsetting in union-mode (with ‘--unn’) causes all variables to be extracted which meet either one or both of the specifications of the variable and group specifications. Union-mode subsetting is simply the logical “OR” of intersection-mode subsetting. As discussed below, the group and variable specifications may be comma separated lists of regular expressions for added control over subsetting.

Remember, if averaging or concatenating large files stresses your systems memory or disk resources, then the easiest solution is often to subset (with ‘-g’ and/or ‘-v’) to retain only the most important variables (see Memory Requirements).

ncks          in.nc out.nc # Extract all groups and variables
ncks -v scl   # Extract variable scl from all groups
ncks -g g1    # Extract group g1 and descendents
ncks -x -g g1 # Extract all groups except g1 and descendents
ncks -g g2,g3 -v scl # Extract scl from groups g2 and g3

Overwriting and appending work as expected:

# Replace scl in group g2 in out.nc with scl from group g2 from in.nc
ncks -A -g g2 -v scl in.nc out.nc

Due to its special capabilities, ncap2 interprets the ‘-v’ switch differently (see ncap2 netCDF Arithmetic Processor). For ncap2, the ‘-v’ switch takes no arguments and indicates that only user-defined variables should be output. ncap2 neither accepts nor understands the -x and -g switches.

Regular expressions the syntax that NCO use pattern-match object names in netCDF file against user requests. The user can select all variables beginning with the string ‘DST’ from an input file by supplying the regular expression ‘^DST’ to the ‘-v’ switch, i.e., ‘-v '^DST'’. The meta-characters used to express pattern matching operations are ‘^$+?.*[]{}|’. If the regular expression pattern matches any part of a variable name then that variable is selected. This capability is also called wildcarding, and is very useful for sub-setting large data files.

Extended regular expressions are defined by the POSIX grep -E (aka egrep) command. As of NCO 2.8.1 (August, 2003), variable name arguments to the ‘-v’ switch may contain extended regular expressions. As of NCO 3.9.6 (January, 2009), variable names arguments to ncatted may contain extended regular expressions. As of NCO 4.2.4 (November, 2012), group name arguments to the ‘-g’ switch may contain extended regular expressions.

Because of its wide availability, NCO uses the POSIX regular expression library regex. Regular expressions of arbitary complexity may be used. Since netCDF variable names are relatively simple constructs, only a few varieties of variable wildcards are likely to be useful. For convenience, we define the most useful pattern matching operators here:

‘^’

Matches the beginning of a string

‘$’

Matches the end of a string

‘.’

Matches any single character

The most useful repetition and combination operators are

‘?’

The preceding regular expression is optional and matched at most once

‘*’

The preceding regular expression will be matched zero or more times

‘+’

The preceding regular expression will be matched one or more times

‘|’

The preceding regular expression will be joined to the following regular expression. The resulting regular expression matches any string matching either subexpression.

To illustrate the use of these operators in extracting variables and groups, consider file in_grp.nc with groups g0–g9, and subgroups s0–s9, in each of those groups, and file in.nc with variables Q, Q01–Q99, Q100, QAA–QZZ, Q_H2O, X_H2O, Q_CO2, X_CO2.

ncks -v '.+' in.nc               # All variables (default)
ncks -v 'Q.?' in.nc              # Variables that contain Q
ncks -v '^Q.?' in.nc             # Variables that start with Q
ncks -v '^Q+.?.' in.nc           # Q, Q0--Q9, Q01--Q99, QAA--QZZ, etc.
ncks -v '^Q..' in.nc             # Q01--Q99, QAA--QZZ, etc.
ncks -v '^Q[0-9][0-9]' in.nc     # Q01--Q99, Q100
ncks -v '^Q[[:digit:]]{2}' in.nc # Q01--Q99
ncks -v 'H2O$' in.nc             # Q_H2O, X_H2O 
ncks -v 'H2O$|CO2$' in.nc        # Q_H2O, X_H2O, Q_CO2, X_CO2 
ncks -v '^Q[0-9][0-9]$' in.nc    # Q01--Q99
ncks -v '^Q[0-6][0-9]|7[0-3]' in.nc # Q01--Q73, Q100
ncks -v '(Q[0-6][0-9]|7[0-3])$' in.nc # Q01--Q73
ncks -v '^[a-z]_[a-z]{3}$' in.nc # Q_H2O, X_H2O, Q_CO2, X_CO2
ncks -g 'g.' in_grp.nc           # 10 Groups g0-g9
ncks -g 's.' in_grp.nc       # 100 sub-groups g0/s0, g0/s1, ... g9/s9
ncks -g 'g.' -v 'v.' in_grp.nc   # All variables 'v.' in groups 'g.'

Beware—two of the most frequently used repetition pattern matching operators, ‘*’ and ‘?’, are also valid pattern matching operators for filename expansion (globbing) at the shell-level. Confusingly, their meanings in extended regular expressions and in shell-level filename expansion are significantly different. In an extended regular expression, ‘*’ matches zero or more occurences of the preceding regular expression. Thus ‘Q*’ selects all variables, and ‘Q+.*’ selects all variables containing ‘Q’ (the ‘+’ ensures the preceding item matches at least once). To match zero or one occurence of the preceding regular expression, use ‘?’. Documentation for the UNIX egrep command details the extended regular expressions which NCO supports.

One must be careful to protect any special characters in the regular expression specification from being interpreted (globbed) by the shell. This is accomplish by enclosing special characters within single or double quotes

ncra -v Q?? in.nc out.nc   # Error: Shell attempts to glob wildcards
ncra -v '^Q+..' in.nc out.nc # Correct: NCO interprets wildcards
ncra -v '^Q+..' in*.nc out.nc # Correct: NCO interprets, Shell globs 

The final example shows that commands may use a combination of variable wildcarding and shell filename expansion (globbing). For globbing, ‘*’ and ‘?’ have nothing to do with the preceding regular expression! In shell-level filename expansion, ‘*’ matches any string, including the null string and ‘?’ matches any single character. Documentation for bash and csh describe the rules of filename expansion (globbing).

3.12 Subsetting Coordinate Variables

By default, coordinates variables associated with any variable appearing in the input-file will be placed in the output-file, even if they are not explicitly specified, e.g., with the ‘-v’ switch. Thus variables with a latitude coordinate lat always carry the values of lat with them into the output-file. This feature can be disabled with ‘-C’, which causes NCO to not automatically add coordinates to the variables appearing in the output-file. However, using ‘-C’ does not preclude the user from including some coordinates in the output files simply by explicitly selecting the coordinates with the -v option. The ‘-c’ option, on the other hand, is a shorthand way of automatically specifying that all coordinate variables in the input-files should appear in the output-file. Thus ‘-c’ allows the user to select all the coordinate variables without having to know their names. As of NCO version 4.4.5 (July, 2014) both ‘-c’ and ‘-C’ honor the CF ancillary_variables convention described in CF Conventions. As of NCO version 4.0.8 (April, 2011) both ‘-c’ and ‘-C’ honor the CF bounds convention described in CF Conventions. As of NCO version 4.4.9 (May, 2015) both ‘-c’ and ‘-C’ honor the CF climatology convention described in CF Conventions. As of NCO version 3.9.6 (January, 2009) both ‘-c’ and ‘-C’ honor the CF coordinates convention described in CF Conventions. As of NCO version 4.6.0 (May, 2016) both ‘-c’ and ‘-C’ honor the CF grid_mapping convention described in CF Conventions.

3.13 Group Path Editing

Group Path Editing, or GPE, allows the user to restructure (i.e., add, remove, and rename groups) in the output file relative to the input file based on the instructions they provide. As of NCO 4.2.3 (November, 2012), all operators that accept netCDF4 files with groups accept the ‘-G’ switch, or its long-option equivalent ‘--gpe’. To master GPE one must understand the meaning of the required gpe_dsc structure/argument that specifies the transformation of input-to-output group paths.

Each gpe_dsc contains up to three elements (two are optional) in the following order:
gpe_dsc = grp_pth:lvl_nbr or grp_pth@lvl_nbr

grp_pth

Group Path. This (optional) component specifies the output group path that should be appended after any editing (i.e., deletion or truncation) of the input path is performed.

lvl_nbr

The number of levels to delete (from the head) or truncate (from the tail) of the input path.

If both components of the argument are present, then a single character, either the colon or at-sign (: or @), must separate them. If only grp_pth is specifed, the separator character may be omitted, e.g., ‘-G g1’. If only lvl_nbr is specifed, the separator character is still required to indicate it is a lvl_nbr arugment and not a grp_pth, e.g., ‘-G :-1’ or ‘-G @1’.

If the at-sign separator character @ is used instead of the colon separator character :, then the following lvl_nbr arugment must be positive and it will be assumed to refer to Truncation-Mode. Hence, ‘-G :-1’ is the same as ‘-G @1’. This is simply a way of making the lvl_nbr argument positive-definite.

3.13.1 Deletion, Truncation, and Flattening of Groups

GPE has three editing modes: Delete, Truncate, and Flatten. Select one of GPE’s three editing modes by supplying a lvl_nbr that is positive, negative, or zero for Delete-, Truncate- and Flatten-mode, respectively.

In Delete-mode, lvl_nbr is a positive integer which specifies the maximum number of group path components (i.e., groups) that GPE will try to delete from the head of grp_pth. For example lvl_nbr = 3 changes the input path /g1/g2/g3/g4/g5 to the output path /g4/g5. Input paths with lvl_nbr or fewer components (groups) are completely erased and the output path commences from the root level.

In other words, GPE is tolerant of specifying too many group components to delete. It deletes as many as possible, without complaint, and then begins to flatten the file (which fails if namespace conflicts arise).

In Truncate-mode, lvl_nbr is a negative integer which specifies the maximum number of group path components (i.e., groups) that GPE will try to truncate from the tail of grp_pth. For example lvl_nbr = -3 changes the input path /g1/g2/g3/g4/g5 to the output path /g1/g2. Input paths with lvl_nbr or fewer components (groups) are completely erased and the output path commences from the root level.

In Flatten-mode, indicated by the separator character alone or with lvl_nbr = 0, GPE removes the entire group path from the input file and constructs the output path beginning at the root level. For example -G :0 and -G : are identical and change the input path /g1/g2/g3/g4/g5 to the output path / whereas -G g1:0 and -G g1: are identical and result in the output path /g1 for all variables.

Subsequent to the alteration of the input path by the specified editing mode, if any, GPE prepends (in Delete Mode) or Appends (in Truncate-mode) any specifed grp_pth to the output path. For example -G g2 changes the input paths / and /g1 to /g2 and /g1/g2, respectively. Likewise, -G g2/g3 changes the input paths / and /g1 to /g2/g3 and /g1/g2/g3, respectively. When grp_pth and lvl_nbr are both specified, the editing actions are taken in sequence so that, e.g., -G g1/g2:2 changes the input paths / and /h1/h2/h3/h4 to /g1/g2 and /g1/g2/h3/h4, respectively. Likewise, -G g1/g2:-2 changes the input paths / and /h1/h2/h3/h4 to /g1/g2 and /h1/h2/g1/g2, respectively.

Combining GPE with subsetting (see Subsetting Files) yields powerful control over the extracted (or excluded) variables and groups and their placement in the output file as shown by the following commands. All commands below may be assumed to end with ‘in.nc out.nc’.

# Prepending paths without editing:
ncks                   # /g?/v? -> /g?/v?
ncks             -v v1 # /g?/v1 -> /g?/v1
ncks       -g g1       # /g1/v? -> /g1/v?
ncks -G o1             # /g?/v? -> /o1/g?/v?
ncks -G o1 -g g1       # /g1/v? -> /o1/g1/v?
ncks       -g g1 -v v1 # /g1/v1 -> /g1/v1
ncks -G o1       -v v1 # /g?/v1 -> /o1/g?/v1
ncks -G o1 -g g1 -v v1 # /g1/v1 -> /o1/g1/v1
ncks -G g1 -g /  -v v1 # /v1    -> /g1/v1
ncks -G g1/g2    -v v1 # /g?/v1 -> /g1/g2/g?/v1
# Delete-mode: Delete from and Prepend to path head
# Syntax: -G [ppn]:lvl_nbr = # of levels to delete
ncks -G :1    -g g1    -v v1 # /g1/v1    -> /v1
ncks -G :1    -g g1/g1 -v v1 # /g1/g1/v1 -> /g1/v1
ncks -G :2    -g g1/g1 -v v1 # /g1/g1/v1 -> /v1
ncks -G :2    -g g1    -v v1 # /g1/v1    -> /v1
ncks -G g2:1  -g g1    -v v1 # /g1/v1    -> /g2/v1
ncks -G g2:2  -g g1/g1 -v v1 # /g1/g1/v1 -> /g2/v1
ncks -G g2:1  -g /     -v v1 # /v1       -> /g2/v1
ncks -G g2:1           -v v1 # /v1       -> /g2/v1
ncks -G g2:1  -g g1/g1 -v v1 # /g1/g1/v1 -> /g2/g1/v1
# Flatten-mode: Remove all input path components
# Syntax: -G [apn]: colon without numerical argument
ncks -G :            -v v1 # /g?/v1    -> /v1
ncks -G :   -g g1    -v v1 # /g1/v1    -> /v1
ncks -G :   -g g1/g1 -v v1 # /g1/g1/v1 -> /v1
ncks -G g2:          -v v1 # /g?/v1    -> /g2/v1
ncks -G g2:                # /g?/v?    -> /g2/v?
ncks -G g2: -g g1/g1 -v v1 # /g1/g1/v1 -> /g2/v1
# Truncate-mode: Truncate from and Append to path tail
# Syntax: -G [apn]:-lvl_nbr = # of levels to truncate
# NB: -G [apn]:-lvl_nbr is equivalent to -G [apn]@lvl_nbr
ncks -G :-1   -g g1    -v v1 # /g1/v1    -> /v1
ncks -G :-1   -g g1/g2 -v v1 # /g1/g2/v1 -> /g1/v1
ncks -G :-2   -g g1/g2 -v v1 # /g1/g2/v1 -> /v1
ncks -G :-2   -g g1    -v v1 # /g1/v1    -> /v1
ncks -G g2:-1          -v v1 # /g?/v1    -> /g2/v1
ncks -G g2:-1 -g g1    -v v1 # /g1/v1    -> /g2/v1
ncks -G g1:-1 -g g1/g2 -v v1 # /g1/g2/v1 -> /g1/g1/v1

3.13.2 Moving Groups

Until fall 2013 (netCDF version 4.3.1-pre1), netCDF contained no library function for renaming groups, and therefore ncrename cannot rename groups. However, NCO built on earlier versions of netCDF than 4.3.1 can use a GPE-based workaround mechanism to “rename” groups. The GPE mechanism actually moves (i.e., copies to a new location) groups, a more arduous procedure than simply renaming them. GPE applies to all selected groups, so, in the general case, one must move only the desired group to a new file, and then merge that new file with the original to obtain a file where the desired group has been “renamed” and all else is unchanged. Here is how to “rename” group /g4 to group /f4 with GPE instead of ncrename

ncks -O -G f4:1 -g g4 ~/nco/data/in_grp.nc ~/tmp.nc # Move /g4 to /f4
ncks -O -x -g g4 ~/nco/data/in_grp.nc ~/out.nc # Excise /g4
ncks -A ~/tmp.nc ~/out.nc # Add /f4 to new file

If the original group g4 is not excised from out.nc (step two above), then the final output file would contain both g4 and a copy named f4. Thus GPE can be used to both “rename” and copy groups. The recommended way to rename groups when when netCDF version 4.3.1 is availale is to use ncrename (see ncrename netCDF Renamer).

One may wish to flatten hierarchical group files for many reasons. These include 1. Obtaining flat netCDF3 files for use with tools that do not work with netCDF4 files, 2. Splitting apart hierarchies to re-assemble into different hierarchies, and 3. Providing a subset of a hierarchical file with the simplest possible storage structure.

ncks -O -G : -g cesm -3 ~/nco/data/cmip5.nc ~/cesm.nc # Extract /cesm to /

The -3 switch ³³ specifies the output dataset should be in netCDF3 format, the -G : option flattens all extracted groups, and the -g cesm option extracts only the cesm group and leaves all other groups (e.g., ecmwf, giss).

3.13.3 Dismembering Files

Let us show how to completely disaggregate (or, more memorably) dismember a hierarchical dataset. For now we take this to mean: store each group as a standalone flat dataset in netCDF3 format. This can be accomplished by looping the previous example over all groups. This script ncdismember dismembers the input file fl_in specified in the first argument and places the resulting files in the directory drc_out specified by the second argument:

cat > ~/ncdismember << 'EOF'
#!/bin/sh

# Purpose: Dismember netCDF4/HDF5 hierarchical files. CF-check them.
# Place each input file group in separate netCDF3 output file
# Described in NCO User Guide at http://nco.sf.net/nco.html#dismember
# Requirements: NCO 4.3.x+, UNIX shell utilities awk, grep, sed
# Optional: Decker CFchecker https://bitbucket.org/mde_/cfchecker

# Usage:
# ncdismember <fl_in> <drc_out> [cf_chk] [cf_vrs] [opt]
# where fl_in is input file/URL to dismember, drc_out is output directory
# CF-compliance check is performed when optional third argument is not '0'
# Default checker is Decker's cfchecker installed locally
# Specify cf_chk=nerc for smallified uploads to NERC checker
# Optional fourth argument cf_vrs is CF version to check
# Optional fifth argument opt passes straight-through to ncks
# Arguments must not use shell expansion/globbing
# NB: ncdismember does not clean-up output directory, so user must
# chmod a+x ~/sh/ncdismember
# Examples:
# ncdismember ~/nco/data/mdl_1.nc /data/zender/tmp
# ncdismember http://dust.ess.uci.edu/nco/mdl_1.nc /tmp
# ncdismember http://thredds-test.ucar.edu/thredds/dodsC/testdods/foo.nc /tmp
# ncdismember ~/nco/data/mdl_1.nc /data/zender/nco/tmp cf
# ncdismember ~/nco/data/mdl_1.nc /data/zender/nco/tmp nerc
# ncdismember ~/nco/data/mdl_1.nc /data/zender/nco/tmp cf 1.3
# ncdismember ~/nco/data/mdl_1.nc /data/zender/nco/tmp cf 1.5 --fix_rec_dmn=all

# Command-line argument defaults
fl_in="${HOME}/nco/data/mdl_1.nc" # [sng] Input file to dismember/check
drc_out="${DATA}/nco/tmp" # [sng] Output directory
cf_chk='0' # [flg] Perform CF-compliance check? Which checker?
cf_vrs='1.5' # [sng] Compliance-check this CF version (e.g., '1.5')
opt='' # [flg] Additional ncks options (e.g., '--fix_rec_dmn=all')
# Use single quotes to pass multiple arguments to opt=${5}
# Otherwise arguments would be seen as ${5}, ${6}, ${7} ...

# Command-line argument option parsing
if [ -n "${1}" ]; then fl_in=${1}; fi
if [ -n "${2}" ]; then drc_out=${2}; fi
if [ -n "${3}" ]; then cf_chk=${3}; fi
if [ -n "${4}" ]; then cf_vrs=${4}; fi
if [ -n "${5}" ]; then opt=${5}; fi

# Prepare output directory
echo "NCO dismembering file ${fl_in}"
fl_stb=$(basename ${fl_in})
drc_out=${drc_out}/${fl_stb}
mkdir -p ${drc_out}
cd ${drc_out}
chk_dck='n'
chk_nrc='n'
if [ ${cf_chk} = 'nerc' ]; then
    chk_nrc='y'
fi # chk_nrc
if [ ${cf_chk} != '0' ] && [ ${cf_chk} != 'nerc' ]; then
    chk_dck='y'
    hash cfchecker 2>/dev/null || { echo >&2 "Local cfchecker command not found, will smallify and upload to NERC checker instead"; chk_nrc='y'; chk_dck='n'; }
fi # !cf_chk
# Obtain group list
grp_lst=`ncks --cdl -m ${fl_in} | grep '// group' | awk '{$1=$2=$3="";sub(/^  */,"",$0);print}'`
IFS=$'\n' # Change Internal-Field-Separator from <Space><Tab><Newline> to <Newline>
for grp_in in ${grp_lst} ; do
    # Replace slashes by dots for output group filenames
    grp_out=`echo ${grp_in} | sed 's/\///' | sed 's/\//./g'`
    if [ "${grp_out}" = '' ]; then grp_out='root' ; fi
    # Tell older NCO/netCDF if HDF4 with --hdf4 switch (signified by .hdf/.HDF suffix)
    hdf4=`echo ${fl_in} | awk '{if(match(tolower($1),".hdf$")) hdf4="--hdf4"; print hdf4}'`
    # Flatten to netCDF3, anchor, no history, no temporary file, padding, HDF4 flag, options
    cmd="ncks -O -3 -G : -g ${grp_in}/ -h --no_tmp_fl --hdr_pad=40 ${hdf4} ${opt} ${fl_in} ${drc_out}/${grp_out}.nc"
    # Use eval in case ${opt} contains multiple arguments separated by whitespace
    eval ${cmd}
    if [ ${chk_dck} = 'y' ]; then
       # Decker checker needs Conventions <= 1.6
       no_bck_sls=`echo ${drc_out}/${grp_out} | sed 's/\\\ / /g'`
       ncatted -h -a Conventions,global,o,c,CF-${cf_vrs} ${no_bck_sls}.nc
    else # !chk_dck
       echo ${drc_out}/${grp_out}.nc
    fi # !chk_dck
done
if [ ${chk_dck} = 'y' ]; then
    echo 'Decker CFchecker reports CF-compliance of each group in flat netCDF3 format'
    cfchecker -c ${cf_vrs} *.nc
fi
if [ ${chk_nrc} = 'y' ]; then
    # Smallification and NERC upload from qdcf script by Phil Rasch (PJR)
    echo 'Using remote CFchecker http://puma.nerc.ac.uk/cgi-bin/cf-checker.pl'
    cf_lcn='http://puma.nerc.ac.uk/cgi-bin/cf-checker.pl'
    for fl in ${drc_out}/*.nc ; do
	fl_sml=${fl}
	cf_out=${fl%.nc}.html
	dmns=`ncdump -h ${fl_in} | sed -n -e '/dimensions/,/variables/p' | grep = | sed -e 's/=.*//'`
	hyp_sml=''
	for dmn in ${dmns}; do
	    dmn_lc=`echo ${dmn} | tr "[:upper:]" "[:lower:]"`
	    if [ ${dmn_lc} = 'lat' ] || [ ${dmn_lc} = 'latitude' ] || [ ${dmn_lc} = 'lon' ] || [ ${dmn_lc} = 'longitude' ] || [ ${dmn_lc} = 'time' ]; then
		hyp_sml=`echo ${hyp_sml}" -d ${dmn},0"`
	    fi # !dmn_lc
	done
	# Create small version of input file by sampling only first element of lat, lon, time
	ncks -O ${hyp_sml} ${fl} ${fl_sml}
	# Send small file to NERC checker
	curl --form cfversion=1.6 --form upload=@${fl_sml} --form press="Check%20file" ${cf_lcn} -o ${cf_out}
	# Strip most HTML to improve readability
	cat ${cf_out} | sed -e "s/<[^>]*>//g" -e "/DOCTYPE/,/\]\]/d" -e "s/CF-Convention//g" -e "s/Output of//g" -e "s/Compliance Checker//g" -e "s/Check another//g" -e "s/CF-Checker follows//g" -e "s/Received//g" -e "s/for NetCDF//g" -e "s/NetCDF format//g" -e "s/against CF version 1//g" -e "s/\.\.\.//g"
	echo "Full NERC compliance-check log for ${fl} in ${cf_out}"
    done
fi # !nerc
EOF
chmod 755 ~/ncdismember # Make command executable
/bin/mv -f ~/ncdismember ~/sh # Store in location on $PATH, e.g., /usr/local/bin

zender@roulee:~$ ncdismember ~/nco/data/mdl_1.nc ${DATA}/nco/tmp
NCO dismembering file /home/zender/nco/data/mdl_1.nc
/data/zender/nco/tmp/mdl_1.nc/cesm.cesm_01.nc
/data/zender/nco/tmp/mdl_1.nc/cesm.cesm_02.nc
/data/zender/nco/tmp/mdl_1.nc/cesm.nc
/data/zender/nco/tmp/mdl_1.nc/ecmwf.ecmwf_01.nc
/data/zender/nco/tmp/mdl_1.nc/ecmwf.ecmwf_02.nc
/data/zender/nco/tmp/mdl_1.nc/ecmwf.nc
/data/zender/nco/tmp/mdl_1.nc/root.nc

A (potentially more portable) binary executable could be written to dismember all groups with a single invocation, yet dismembering without loss of information is possible now with this simple script on all platforms with UNIXy utilities. Note that all dimensions inherited by groups in the input file are correctly placed by ncdismember into the flat files. Moreover, each output file preserves the group metadata of all ancestor groups, including the global metadata from the input file. As written, the script could fail on groups that contain advanced netCDF4 features because the user requests (with the ‘-3’ switch) that output be netCDF3 classic format. However, ncks detects many format incompatibilities in advance and works around them. For example, ncks autoconverts netCDF4-only atomic-types (such as NC_STRING and NC_UBYTE) to corresponding netCDF3 atomic types (NC_CHAR and NC_SHORT) when the output format is netCDF3.

3.13.4 Checking CF-compliance

One application of dismembering is to check the CF-compliance of each group in a file. When invoked with the optional third argumnt ‘cf’, ncdismember passes each file it generates to freely available compliance checkers, such as cfchecker ³⁴.

zender@roulee:~$ ncdismember ~/nco/data/mdl_1.nc /data/zender/nco/tmp cf
NCO dismembering file /home/zender/nco/data/mdl_1.nc
CFchecker reports CF-compliance of each group in flat netCDF3 format
WARNING: Using the default (non-CF) Udunits database
cesm.cesm_01.nc: 
INFO: INIT:     running CFchecker version 1.5.15
INFO: INIT:     checking compliance with convention CF-1.5
INFO: INIT:     using standard name table version: 25, last modified: 2013-07-05T05:40:30Z
INFO: INIT:     using area type table version: 2, date: 10 July 2013
INFO: 2.4:      no axis information found in dimension variables, not checking dimension order
WARNING: 3:     variable "tas1" contains neither long_name nor standard_name attribute
WARNING: 3:     variable "tas2" contains neither long_name nor standard_name attribute
INFO: 3.1:      variable "tas1" does not contain units attribute
INFO: 3.1:      variable "tas2" does not contain units attribute
--------------------------------------------------
cesm.cesm_02.nc: 
...

By default the CF version checked is determined automatically by cfchecker. The user can override this default by supplying a supported CF version, e.g., ‘1.3’, as an optional fourth argument to ncdismember. Current valid CF options are ‘1.0’, ‘1.1’, ‘1.2’, ‘1.3’, ‘1.4’, and ‘1.5’.

Our development and testing of ncdismember is funded by our involvement in NASA’s Dataset Interoperability Working Group (DIWG), though our interest extends beyond NASA datasets. Taken together, NCO’s features (autoconversion to netCDF3 atomic types, fixing multiple record dimensions, autosensing HDF4 input, scoping rules for CF conventions) make ncdismember reliable and friendly for both dismembering hierarchical files and for CF-compliance checks. Most HDF4 and HDF5 datasets can be checked for CF-compliance with a one-line command. Example compliance checks of common NASA datasets are at http://dust.ess.uci.edu/diwg. Our long-term goal is to enrich the hierarchical data model with the expressivity and syntactic power of CF conventions.

NASA asked the DIWG to prepare a one-page summary of the procedure necessary to check HDF files for CF-compliance:

cat > ~/ncdismember.txt << 'EOF'
    Preparing an RPM-based OS to Test HDF & netCDF Files for CF-Compliance

By Charlie Zender, UCI & NASA Dataset Interoperability Working Group (DIWG)

Installation Summary:
1. HDF4 [with internal netCDF support _disabled_]
2. HDF5
3. netCDF [with external HDF4 support _enabled_]
4. NCO
5. numpy
6. netcdf4-python
7. python-lxml
8. CFunits-python
9. CFChecker
10. ncdismember

All 10 packages can use default installs _except_ HDF4 and netCDF.
Following instructions for Fedora Core 20 (FC20), an RPM-based Linux OS
Feedback and changes for other Linux-based OS's welcome to zender at uci.edu
${H4DIR}, ${H5DIR}, ${NETCDFDIR}, ${NCODIR}, may all be different
For simplicity CZ sets them all to /usr/local

# 1. HDF4. Build in non-default manner. Turn-off its own netCDF support.
# Per http://www.unidata.ucar.edu/software/netcdf/docs/build_hdf4.html
# HDF4 support not necessary though it makes ncdismember more comprehensive
wget -c http://www.hdfgroup.org/ftp/HDF/HDF_Current/src/hdf-4.2.9.tar.gz
tar xvzf hdf-4.2.9.tar.gz
cd hdf-4.2.9
./configure --enable-shared --disable-netcdf --disable-fortran --prefix=${H4DIR}
make && make check && make install

# 2. HDF5. Build normally. RPM may work too. Please let me know if so.
# HDF5 is a necessary pre-requisite for netCDF4
wget -c ftp://ftp.unidata.ucar.edu/pub/netcdf/netcdf-4/hdf5-1.8.11.tar.gz
tar xvzf hdf5-1.8.11.tar.gz
cd hdf5-1.8.11
./configure --enable-shared --prefix=${H5DIR}
make && make check && make install

# 3. netCDF version 4.3.1 or later. Build in non-default manner with HDF4.
# Per http://www.unidata.ucar.edu/software/netcdf/docs/build_hdf4.html
# Earlier versions of netCDF may fail checking some HDF4 files
wget -c ftp://ftp.unidata.ucar.edu/pub/netcdf/netcdf-4.3.2.tar.gz
tar xvzf netcdf-4.3.2.tar.gz
cd netcdf-4.3.2
CPPFLAGS="-I${H5DIR}/include -I${H4DIR}/include" \
LDFLAGS="-L${H5DIR}/lib -L${H4DIR}/lib" \
./configure --enable-hdf4 --enable-hdf4-file-tests
make && make check && make install

# 4. NCO version 4.4.0 or later. Some RPMs available. Or install by hand.
# Later versions of NCO have much better support for ncdismember
wget http://nco.sourceforge.net/src/nco-4.4.4.tar.gz .
tar xvzf nco-4.4.4.tar.gz
cd nco-4.4.4
./configure --prefix=${NCODIR}
make && make install

# 5. numpy
sudo yum install numpy -y

# 6. netcdf4-python
sudo yum install netcdf4-python -y

# 7. python-lxml
sudo yum install python-lxml -y

# 8. CFunits-python. No RPM available. Must install by hand.
# http://code.google.com/p/cfunits-python/
wget http://cfunits-python.googlecode.com/files/cfunits-0.9.6.tar.gz .
tar xvzf cfunits-0.9.6.tar.gz
cd cfunits-0.9.6
sudo python setup.py install

# 9. CFChecker. No RPM available. Must install by hand.
# https://bitbucket.org/mde_/cfchecker
wget https://bitbucket.org/mde_/cfchecker/downloads/CFchecker-1.5.15.tar.bz2 . 
tar xvjf CFchecker-1.5.15.tar.bz2 
cd CFchecker
sudo python setup.py install

# 10. ncdismember. Copy script from http://nco.sf.net/nco.html#ncdismember
# Store dismembered files somewhere, e.g., ${DATA}/nco/tmp/hdf
mkdir -p ${DATA}/nco/tmp/hdf
# Many datasets work with a simpler command...
ncdismember ~/nco/data/in.nc ${DATA}/nco/tmp/hdf cf 1.5
ncdismember ~/nco/data/mdl_1.nc ${DATA}/nco/tmp/hdf cf 1.5
ncdismember ${DATA}/hdf/AMSR_E_L2_Rain_V10_200905312326_A.hdf \
            ${DATA}/nco/tmp/hdf cf 1.5
ncdismember ${DATA}/hdf/BUV-Nimbus04_L3zm_v01-00-2012m0203t144121.h5 \
            ${DATA}/nco/tmp/hdf cf 1.5
ncdismember ${DATA}/hdf/HIRDLS-Aura_L3ZAD_v06-00-00-c02_2005d022-2008d077.he5 ${DATA}/nco/tmp/hdf cf 1.5
# Some datasets, typically .h5, require the --fix_rec_dmn=all argument
ncdismember_${DATA}/hdf/GATMO_npp_d20100906_t1935191_e1935505_b00012_c20110707155932065809_noaa_ops.h5 ${DATA}/nco/tmp/hdf cf 1.5 --fix_rec_dmn=all
ncdismember ${DATA}/hdf/mabel_l2_20130927t201800_008_1.h5 \
            ${DATA}/nco/tmp/hdf cf 1.5 --fix_rec_dmn=all
EOF

A PDF version of these instructions is available here.

3.14 C and Fortran Index conventions

The ‘-F’ switch changes NCO to read and write with the Fortran index convention. By default, NCO uses C-style (0-based) indices for all I/O. In C, indices count from 0 (rather than 1), and dimensions are ordered from slowest (inner-most) to fastest (outer-most) varying. In Fortran, indices count from 1 (rather than 0), and dimensions are ordered from fastest (inner-most) to slowest (outer-most) varying. Hence C and Fortran data storage conventions represent mathematical transposes of eachother. Note that record variables contain the record dimension as the most slowly varying dimension. See ncpdq netCDF Permute Dimensions Quickly for techniques to re-order (including transpose) dimensions and to reverse data storage order.

Consider a file 85.nc containing 12 months of data in the record dimension time. The following hyperslab operations produce identical results, a June-July-August average of the data:

ncra -d time,5,7 85.nc 85_JJA.nc
ncra -F -d time,6,8 85.nc 85_JJA.nc

Printing variable three_dmn_var in file in.nc first with the C indexing convention, then with Fortran indexing convention results in the following output formats:

% ncks -v three_dmn_var in.nc
lat[0]=-90 lev[0]=1000 lon[0]=-180 three_dmn_var[0]=0 
...
% ncks -F -v three_dmn_var in.nc
lon(1)=0 lev(1)=100 lat(1)=-90 three_dmn_var(1)=0 
...

3.15 Hyperslabs

A hyperslab is a subset of a variable’s data. The coordinates of a hyperslab are specified with the -d dim,[min][,[max][,[stride]]] short option (or with the same arguments to the ‘--dimension’ or ‘--dmn’ long options). At least one hyperslab argument (min, max, or stride) must be present. The bounds of the hyperslab to be extracted are specified by the associated min and max values. A half-open range is specified by omitting either the min or max parameter. The separating comma must be present to indicate the omission of one of these arguments. The unspecified limit is interpreted as the maximum or minimum value in the unspecified direction. A cross-section at a specific coordinate is extracted by specifying only the min limit and omitting a trailing comma. Dimensions not mentioned are passed with no reduction in range. The dimensionality of variables is not reduced (in the case of a cross-section, the size of the constant dimension will be one).

# First and second longitudes
ncks -F -d lon,1,2 in.nc out.nc
# Second and third longitudes
ncks -d lon,1,2 in.nc out.nc

As of version 4.2.1 (August, 2012), NCO allows one to extract the last N elements of a hyperslab. Negative integers as min or max elements of a hyperslab specification indicate offsets from the end (Python also uses this convention). Consistent with this convention, the value ‘-1’ (negative one) indicates the last element of a dimension, and negative zero is algebraically equivalent to zero and so indicates the first element of a dimension. Previously, for example, ‘-d time,-2,-1’ caused a domain error. Now it means select the penultimate and last timesteps, independent of the size of the time dimension. Select only the first and last timesteps, respectively, with ‘-d time,0’ and ‘-d time,-1’. Negative integers work for min and max indices, though not for stride.

# Second through penultimate longitudes
ncks -d lon,1,-2 in.nc out.nc
# Second through last longitude
ncks -d lon,1,-1 in.nc out.nc
# Second-to-last to last longitude
ncks -d lon,-3,-1 in.nc out.nc
# Second-to-last to last longitude 
ncks -d lon,-3, in.nc out.nc

The ‘-F’ argument, if any, applies the Fortran index convention only to indices specified as positive integers:

# First through penultimate longitudes
ncks -F -d lon,1,-2 in.nc out.nc (-F affects only start index)
# First through last longitude
ncks -F -d lon,1,-1 in.nc out.nc
# Second-to-last to penultimate longitude (-F has no effect)
ncks -F -d lon,-3,-1 in.nc out.nc
# Second-to-last to last longitude (-F has no effect)
ncks -F -d lon,-3, in.nc out.nc

Coordinate values should be specified using real notation with a decimal point required in the value, whereas dimension indices are specified using integer notation without a decimal point. This convention serves only to differentiate coordinate values from dimension indices. It is independent of the type of any netCDF coordinate variables. For a given dimension, the specified limits must both be coordinate values (with decimal points) or dimension indices (no decimal points).

If values of a coordinate-variable are used to specify a range or cross-section, then the coordinate variable must be monotonic (values either increasing or decreasing). In this case, command-line values need not exactly match coordinate values for the specified dimension. Ranges are determined by seeking the first coordinate value to occur in the closed range [min,max] and including all subsequent values until one falls outside the range. The coordinate value for a cross-section is the coordinate-variable value closest to the specified value and must lie within the range or coordinate-variable values. The stride argument, if any, must be a dimension index, not a coordinate value. See Stride, for more information on the stride option.

# All longitude values between 1 and 2 degrees
ncks -d lon,1.0,2.0 in.nc out.nc
# All longitude values between 1 and 2 degrees
ncks -F -d lon,1.0,2.0 in.nc out.nc
# Every other longitude value between 0 and 90 degrees
ncks -F -d lon,0.0,90.0,2 in.nc out.nc

As shown, we recommend using a full floating-point suffix of .0 instead of simply . in order to make obvious the selection of hyperslab elements based on coordinate value rather than index.

User-specified coordinate limits are promoted to double-precision values while searching for the indices which bracket the range. Thus, hyperslabs on coordinates of type NC_CHAR are computed numerically rather than lexically, so the results are unpredictable.

The relative magnitude of min and max indicate to the operator whether to expect a wrapped coordinate (see Wrapped Coordinates), such as longitude. If min > max, the NCO expects the coordinate to be wrapped, and a warning message will be printed. When this occurs, NCO selects all values outside the domain [max < min], i.e., all the values exclusive of the values which would have been selected if min and max were swapped. If this seems confusing, test your command on just the coordinate variables with ncks, and then examine the output to ensure NCO selected the hyperslab you expected (coordinate wrapping is currently only supported by ncks).

Because of the way wrapped coordinates are interpreted, it is very important to make sure you always specify hyperslabs in the monotonically increasing sense, i.e., min < max (even if the underlying coordinate variable is monotonically decreasing). The only exception to this is when you are indeed specifying a wrapped coordinate. The distinction is crucial to understand because the points selected by, e.g., -d longitude,50.,340., are exactly the complement of the points selected by -d longitude,340.,50..

Not specifying any hyperslab option is equivalent to specifying full ranges of all dimensions. This option may be specified more than once in a single command (each hyperslabbed dimension requires its own -d option).

3.16 Stride

All data operators support specifying a stride for any and all dimensions at the same time. The stride is the spacing between consecutive points in a hyperslab. A stride of 1 picks all the elements of the hyperslab, and a stride of 2 skips every other element, etc.. ncks multislabs support strides, and are more powerful than the regular hyperslabs supported by the other operators (see Multislabs). Using the stride option for the record dimension with ncra and ncrcat makes it possible, for instance, to average or concatenate regular intervals across multi-file input data sets.

The stride is specified as the optional fourth argument to the ‘-d’ hyperslab specification: -d dim,[min][,[max][,[stride]]]. Specify stride as an integer (i.e., no decimal point) following the third comma in the ‘-d’ argument. There is no default value for stride. Thus using ‘-d time,,,2’ is valid but ‘-d time,,,2.0’ and ‘-d time,,,’ are not. When stride is specified but min is not, there is an ambiguity as to whether the extracted hyperslab should begin with (using C-style, 0-based indexes) element 0 or element ‘stride-1’. NCO must resolve this ambiguity and it chooses element 0 as the first element of the hyperslab when min is not specified. Thus ‘-d time,,,stride’ is syntactically equivalent to ‘-d time,0,,stride’. This means, for example, that specifying the operation ‘-d time,,,2’ on the array ‘1,2,3,4,5’ selects the hyperslab ‘1,3,5’. To obtain the hyperslab ‘2,4’ instead, simply explicitly specify the starting index as 1, i.e., ‘-d time,1,,2’.

For example, consider a file 8501_8912.nc which contains 60 consecutive months of data. Say you wish to obtain just the March data from this file. Using 0-based subscripts (see C and Fortran Index Conventions) these data are stored in records 2, 14, … 50 so the desired stride is 12. Without the stride option, the procedure is very awkward. One could use ncks five times and then use ncrcat to concatenate the resulting files together:

for idx in 02 14 26 38 50; do # Bourne Shell
  ncks -d time,${idx} 8501_8912.nc foo.${idx}
done
foreach idx (02 14 26 38 50) # C Shell
  ncks -d time,${idx} 8501_8912.nc foo.${idx}
end
ncrcat foo.?? 8589_03.nc
rm foo.??

With the stride option, ncks performs this hyperslab extraction in one operation:

ncks -d time,2,,12 8501_8912.nc 8589_03.nc

See ncks netCDF Kitchen Sink, for more information on ncks.

Applying the stride option to the record dimension in ncra and ncrcat makes it possible, for instance, to average or concatenate regular intervals across multi-file input data sets.

ncra -F -d time,3,,12 85.nc 86.nc 87.nc 88.nc 89.nc 8589_03.nc
ncrcat -F -d time,3,,12 85.nc 86.nc 87.nc 88.nc 89.nc 8503_8903.nc

3.17 Record Appending

As of version 4.2.6 (March, 2013), NCO allows both Multi-File, Multi-Record operators (ncra and ncrcat) to append their output directly to the end of an existing file. This feature may be used to augment a target file, rather than construct it from scratch. This helps, for example, when a timeseries is concatenated from input data that becomes available in stages rather than all at once. In such cases this switch significantly speeds writing.

Consider the use case where one wishes to preserve the contents of fl_1.nc, and add to them new records contained in fl_2.nc. Previously the output had to be placed in a third file, fl_3.nc (which could also safely be named fl_2.nc), via

ncrcat -O fl_1.nc fl_2.nc fl_3.nc

Under the hood this operation copies all information in fl_1.nc and fl_2.nc not once but twice. The first copy is performed through the netCDF interface, as all data from fl_1.nc and fl_2.nc are extracted and placed in the output file. The second copy occurs (usually much) more quickly as the (by default) temporary output file is copied (sometimes a quick re-link suffices) to the final output file (see Temporary Output Files). All this copying is expensive for large files.

The ‘--record_append’ switch appends all records in fl_2.nc to the end (after the last record) of fl_1.nc:

ncrcat --rec_apn fl_2.nc fl_1.nc

The ordering of the filename arguments may seem non-intuitive. If the record variable represents time in these files, then the values in fl_1.nc precede those in fl_2.nc, so why do the files appear in the reverse order on the command line? fl_1.nc is the last file named because it is the pre-existing output file to which we will append all the other input files listed (in this case only fl_2.nc). The contents of fl_1.nc are completely preserved, and only values in fl_2.nc (and any other input files) are copied. This switch avoids the necessity of copying all of fl_1.nc through the netCDF interface to a new output file. The ‘--rec_apn’ switch automatically puts NCO into append mode (see Appending Variables), so specifying ‘-A’ is redundant, and simultaneously specifying overwrite mode with ‘-O’ causes an error. By default, NCO works in an intermediate temporary file. Power users may combine ‘--rec_apn’ with the ‘--no_tmp_fl’ switch (see Temporary Output Files):

ncrcat --rec_apn --no_tmp_fl fl_2.nc fl_1.nc

This avoids creating an intermediate file, and copies only the minimal amount of data (i.e., all of fl_2.nc). Hence, it is fast. We recommend users try to understand the safety trade-offs involved.

3.18 Subcycle

As of version 4.2.1 (August, 2012), NCO allows both Multi-File, Multi-Record operators, ncra and ncrcat, to extract and operate on multiple groups of records. These groups may be connected to physical sub-cycles of a periodic nature, e.g., months of a year, or hours of a day. Or they may be thought of as groups of a specifed duration. The feature and the terminology to describe it are new. For now, we call this the subcycle feature, sometimes abbreviated SSC ³⁵.

The subcycle feature allows processing of groups of records separated by regular intervals of records. It is perhaps best illustrated by an extended example which describes how to solve the same problem both with and without the SSC feature.

The first task in climate data processing is often creating seasonal cycles. Suppose a 150-year climate simulation produces 150 output files, each comprising 12 records, each record a monthly mean: 1850.nc, 1851.nc, ... 1999.nc. Our goal is to create a single file containing the summertime (June, July, and August, aka JJA) mean. Traditionally, we would first compute the climatological monthly mean for each month of summer. Each of these is a 150-year mean, i.e.,

# Step 1: Create climatological monthly files clm06.nc..clm08.nc
for mth in {6..8}; do
  mm=`printf "%02d" $mth`
  ncra -O -F -d time,${mm},,12 -n 150,4,1 1850.nc clm${mm}.nc
done
# Step 2: Average climatological monthly files into summertime mean
ncra -O clm06 clm07.nc clm08.nc clm_JJA.nc

So far, nothing is unusual and this task can be performed by any NCO version. The SSC feature makes obsolete the need for the shell loop used in Step 1 above.

The new SSC option aggregates more than one input record at a time before performing arithmetic operations, and, with an additional switch, allows us to archive those results in multiple-record output (MRO) files. This reduces the task of producing the climatological summertime mean to one step:

# Step 1: Compute climatological summertime mean
ncra -O -F -d time,6,,12,3 -n 150,4,1 1850.nc clm_JJA.nc

The SSC option instructs ncra (or ncrcat) to process files in groups of three records. To better understand the meaning of each argument to the ‘-d’ hyperslab option, read it this way: “for the time dimension start with the sixth record, continue without end, repeat the process every twelfth record, and define a sub-cycle as three consecutive records”.

A separate option, ‘--mro’, instructs ncra to output its results from each sub-group, and to produce a Multi-Record Output (MRO) file rather than a Single-Record Output (SRO) file. Unless ‘--mro’ is specified, ncra collects together all the sub-groups, operates on their ensemble, and produces a single output record. The addition of ‘--mro’ to the above example causes ncra to archive all (150) annual summertime means to one file:

# Step 1: Archive all 150 summertime means in one file
ncra --mro -O -F -d time,6,,12,3 -n 150,4,1 1850.nc 1850_2009_JJA.nc
# ...or all (150) annual means...
ncra --mro -O -d time,,,12,12 -n 150,4,1 1850.nc 1850_2009.nc

These operations generate and require no intermediate files. This contrasts to previous NCO methods, which require generating, averaging, then catenating 150 files. The ‘--mro’ option has no effect on, or rather is redundant for, ncrcat since ncrcat always outputs all selected records.

3.19 Multislabs

A multislab is a union of one or more hyperslabs. One defines multislabs by chaining together hyperslab commands, i.e., -d options (see Hyperslabs). Support for specifying a multi-hyperslab or multislab for any variable was first added to ncks in late 2002. The other operators received these capabilities in April 2008. Multi-slabbing is often referred to by the acronym MSA, which stands for “Multi-Slabbing Algorithm”. As explained below, the user may additionally request that the multislabs be returned in the user-specified order, rather than the on-disk storage order. Although MSA user-ordering has been available in all operators since 2008, most users were unaware of it since the documentation (below, and in the man pages) was not written until July 2013.

Multislabs overcome many restraints that limit simple hyperslabs. A single -d option can only specify a contiguous and/or a regularly spaced multi-dimensional data array. Multislabs are constructed from multiple -d options and may therefore have non-regularly spaced arrays. For example, suppose it is desired to operate on all longitudes from 10.0 to 20.0 and from 80.0 to 90.0 degrees. The combined range of longitudes is not selectable in a single hyperslab specfication of the form ‘-d dimension,min,max’ or ‘-d dimension,min,max,stride’ because its elements are irregularly spaced in coordinate space (and presumably in index space too). The multislab specification for obtaining these values is simply the union of the hyperslabs specifications that comprise the multislab, i.e.,

ncks -d lon,10.,20. -d lon,80.,90. in.nc out.nc
ncks -d lon,10.,15. -d lon,15.,20. -d lon,80.,90. in.nc out.nc

Any number of hyperslabs specifications may be chained together to specify the multislab. MSA creates an output dimension equal in size to the sum of the sizes of the multislabs. This can be used to extend and or pad coordinate grids.

Users may specify redundant ranges of indices in a multislab, e.g.,

ncks -d lon,0,4 -d lon,2,9,2 in.nc out.nc

This command retrieves the first five longitudes, and then every other longitude value up to the tenth. Elements 0, 2, and 4 are specified by both hyperslab arguments (hence this is redundant) but will count only once if an arithmetic operation is being performed. This example uses index-based (not coordinate-based) multislabs because the stride option only supports index-based hyper-slabbing. See Stride, for more information on the stride option.

Multislabs are more efficient than the alternative of sequentially performing hyperslab operations and concatenating the results. This is because NCO employs a novel multislab algorithm to minimize the number of I/O operations when retrieving irregularly spaced data from disk. The NCO multislab algorithm retrieves each element from disk once and only once. Thus users may take some shortcuts in specifying multislabs and the algorithm will obtain the intended values. Specifying redundant ranges is not encouraged, but may be useful on occasion and will not result in unintended consequences.

Suppose the Q variable contains three dimensional arrays of distinct chemical constituents in no particular order. We are interested in the NOy species in a certain geographic range. Say that NO, NO2, and N2O5 are elements 0, 1, and 5 of the species dimension of Q. The multislab specification might look something like

ncks -d species,0,1 -d species,5 -d lon,0,4 -d lon,2,9,2 in.nc out.nc

Multislabs are powerful because they may be specified for every dimension at the same time. Thus multislabs obsolete the need to execute multiple ncks commands to gather the desired range of data.

The MSA user-order switch ‘--msa_usr_rdr’ (or ‘--msa_user_order’, both of which shorten to ‘--msa’) requests that the multislabs be output in the user-specified order from the command-line, rather than in the input-file on-disk storage order. This allows the user to perform complex data re-ordering in one operation that would otherwise require cumbersome steps of hyperslabbing, concatenating, and permuting. Consider the recent example of a user who needed to convert datasets stored with the longitude coordinate Lon ranging from [-180,180) to datasets that follow the [0,360) convention.

% ncks -H -v Lon in.nc
Lon[0]=-180
Lon[1]=-90
Lon[2]=0
Lon[3]=90

Although simple in theory, this task requires both mathematics to change the numerical value of the longitude coordinate, data hyperslabbing to split the input on-disk arrays at Greenwich, and data re-ordering within to stitch the western hemisphere onto the eastern hemisphere at the date-line. The ‘--msa’ user-order switch overrides the default that data are output in the same order in which they are stored on-disk in the input file, and instead stores them in the same order as the multi-slabs are given to the command line. This default is intuitive and is not important in most uses. However, the MSA user-order switch allows users to meet their output order needs by specifying multi-slabs in a certain order. Compare the results of default ordering to user-ordering for longitude:

% ncks -O -H       -v Lon -d Lon,0.,180. -d Lon,-180.,-1.0 in.nc
Lon[0]=-180 
Lon[1]=-90 
Lon[2]=0 
Lon[3]=90 
% ncks -O -H --msa -v Lon -d Lon,0.,180. -d Lon,-180.,-1.0 in.nc
Lon[0]=0 
Lon[1]=90 
Lon[2]=-180 
Lon[3]=-90 

The two multi-slabs are the same but they can be presented to screen, or to an output file, in either order. The second example shows how to place the western hemisphere after the eastern hemisphere, although they are stored in the opposite order in the input file.

With this background, one sees that the following commands suffice to rotate the input file by 180 degrees longitude:

% ncks -O -v LatLon --msa -d Lon,0.,180. -d Lon,-180.,-1.0 in.nc out.nc
% ncap2 -O -s 'where(Lon < 0) Lon=Lon+360' out.nc out.nc
% ncks -C -H -v LatLon ~/nco/data/in.nc
Lat[0]=-45 Lon[0]=-180 LatLon[0]=0 
Lat[0]=-45 Lon[1]=-90 LatLon[1]=1 
Lat[0]=-45 Lon[2]=0 LatLon[2]=2 
Lat[0]=-45 Lon[3]=90 LatLon[3]=3 
Lat[1]=45 Lon[0]=-180 LatLon[4]=4 
Lat[1]=45 Lon[1]=-90 LatLon[5]=5 
Lat[1]=45 Lon[2]=0 LatLon[6]=6 
Lat[1]=45 Lon[3]=90 LatLon[7]=7 
% ncks -C -H -v LatLon ~/out.nc
Lat[0]=-45 Lon[0]=0 LatLon[0]=2 
Lat[0]=-45 Lon[1]=90 LatLon[1]=3 
Lat[0]=-45 Lon[2]=180 LatLon[2]=0 
Lat[0]=-45 Lon[3]=270 LatLon[3]=1 
Lat[1]=45 Lon[0]=0 LatLon[4]=6 
Lat[1]=45 Lon[1]=90 LatLon[5]=7 
Lat[1]=45 Lon[2]=180 LatLon[6]=4 
Lat[1]=45 Lon[3]=270 LatLon[7]=5 

There are other workable, valid methods to accomplish this rotation, yet none are simpler nor more efficient than utilizing MSA user-ordering. Some final comments on applying this algorithm: Be careful to specify hemispheres that do not overlap, e.g., by inadvertently specifying coordinate ranges that both include Greenwich. Some users will find using index-based rather than coordinate-based hyperslabs makes this clearer.

3.20 Wrapped Coordinates

A wrapped coordinate is a coordinate whose values increase or decrease monotonically (nothing unusual so far), but which represents a dimension that ends where it begins (i.e., wraps around on itself). Longitude (i.e., degrees on a circle) is a familiar example of a wrapped coordinate. Longitude increases to the East of Greenwich, England, where it is defined to be zero. Halfway around the globe, the longitude is 180 degrees East (or West). Continuing eastward, longitude increases to 360 degrees East at Greenwich. The longitude values of most geophysical data are either in the range [0,360), or [-180,180). In either case, the Westernmost and Easternmost longitudes are numerically separated by 360 degrees, but represent contiguous regions on the globe. For example, the Saharan desert stretches from roughly 340 to 50 degrees East. Extracting the hyperslab of data representing the Sahara from a global dataset presents special problems when the global dataset is stored consecutively in longitude from 0 to 360 degrees. This is because the data for the Sahara will not be contiguous in the input-file but is expected by the user to be contiguous in the output-file. In this case, ncks must invoke special software routines to assemble the desired output hyperslab from multiple reads of the input-file.

Assume the domain of the monotonically increasing longitude coordinate lon is 0 < lon < 360. ncks will extract a hyperslab which crosses the Greenwich meridian simply by specifying the westernmost longitude as min and the easternmost longitude as max. The following commands extract a hyperslab containing the Saharan desert:

ncks -d lon,340.,50. in.nc out.nc
ncks -d lon,340.,50. -d lat,10.,35. in.nc out.nc

The first example selects data in the same longitude range as the Sahara. The second example further constrains the data to having the same latitude as the Sahara. The coordinate lon in the output-file, out.nc, will no longer be monotonic! The values of lon will be, e.g., ‘340, 350, 0, 10, 20, 30, 40, 50’. This can have serious implications should you run out.nc through another operation which expects the lon coordinate to be monotonically increasing. Fortunately, the chances of this happening are slim, since lon has already been hyperslabbed, there should be no reason to hyperslab lon again. Should you need to hyperslab lon again, be sure to give dimensional indices as the hyperslab arguments, rather than coordinate values (see Hyperslabs).

3.21 Auxiliary Coordinates

Utilize auxiliary coordinates specified in values of the coordinate variable’s standard_name attributes, if any, when interpreting hyperslab and multi-slab options. Also ‘--auxiliary’. This switch supports hyperslabbing cell-based grids (aka unstructured grids) over coordinate ranges. This works on datasets that associate coordinate variables to grid-mappings using the CF-convention (see CF Conventions) coordinates and standard_name attributes described here. Currently, NCO understands auxiliary coordinate variables pointed to by the standard_name attributes for latitude and longitude. Cells that contain a value within the user-specified range [lon_min,lon_max,lat_min,lat_max] are included in the output hyperslab.

A cell-based or unstructured grid collapses the horizontal spatial information (latitude and longitude) and stores it along a one-dimensional coordinate that has a one-to-one mapping to both latitude and longitude coordinates. Rectangular (in longitude and latitude) horizontal hyperslabs cannot be selected using the typical procedure (see Hyperslabs) of separately specifying ‘-d’ arguments for longitude and latitude. Instead, when the ‘-X’ is used, NCO learns the names of the latitude and longitude coordinates by searching the standard_name attribute of all variables until it finds the two variables whose standard_name’s are “latitude” and “longitude”, respectively. This standard_name attribute for latitude and longitude coordinates follows the CF-convention (see CF Conventions).

Putting it all together, consider a variable gds_3dvar output from simulations on a cell-based geodesic grid. Although the variable contains three dimensions of data (time, latitude, and longitude), it is stored in the netCDF file with only two dimensions, time and gds_crd.

% ncks -m -C -v gds_3dvar ~/nco/data/in.nc
gds_3dvar: type NC_FLOAT, 2 dimensions, 4 attributes, chunked? no, \
 compressed? no, packed? no, ID = 41
gds_3dvar RAM size is 10*8*sizeof(NC_FLOAT) = 80*4 = 320 bytes
gds_3dvar dimension 0: time, size = 10 NC_DOUBLE, dim. ID = 20 \ 
 (CRD)(REC)
gds_3dvar dimension 1: gds_crd, size = 8 NC_FLOAT, dim. ID = 17 (CRD)
gds_3dvar attribute 0: long_name, size = 17 NC_CHAR, value = \ 
 Geodesic variable
gds_3dvar attribute 1: units, size = 5 NC_CHAR, value = meter
gds_3dvar attribute 2: coordinates, size = 15 NC_CHAR, value = \
 lat_gds lon_gds
gds_3dvar attribute 3: purpose, size = 64 NC_CHAR, value = \ 
 Test auxiliary coordinates like those that define geodesic grids

The coordinates attribute lists the names of the latitude and longitude coordinates, lat_gds and lon_gds, respectively. The coordinates attribute is recommended though optional. With it, the user can immediately identify which variables contain the latitude and longitude coordinates. Without a coordinates attribute it would be unclear at first glance whether a variable resides on a cell-based grid. In this example, time is a normal record dimension and gds_crd is the cell-based dimension.

The cell-based grid file must contain two variables whose standard_name attributes are “latitude”, and “longitude”:

% ncks -m -C -v lat_gds,lon_gds ~/nco/data/in.nc
lat_gds: type NC_DOUBLE, 1 dimensions, 4 attributes, \
 chunked? no, compressed? no, packed? no, ID = 37
lat_gds RAM size is 8*sizeof(NC_DOUBLE) = 8*8 = 64 bytes
lat_gds dimension 0: gds_crd, size = 8 NC_FLOAT, dim. ID = 17 (CRD)
lat_gds attribute 0: long_name, size = 8 NC_CHAR, value = Latitude
lat_gds attribute 1: standard_name, size = 8 NC_CHAR, value = latitude
lat_gds attribute 2: units, size = 6 NC_CHAR, value = degree
lat_gds attribute 3: purpose, size = 62 NC_CHAR, value = \ 
 1-D latitude coordinate referred to by geodesic grid variables

lon_gds: type NC_DOUBLE, 1 dimensions, 4 attributes, \
 chunked? no, compressed? no, packed? no, ID = 38
lon_gds RAM size is 8*sizeof(NC_DOUBLE) = 8*8 = 64 bytes
lon_gds dimension 0: gds_crd, size = 8 NC_FLOAT, dim. ID = 17 (CRD)
lon_gds attribute 0: long_name, size = 9 NC_CHAR, value = Longitude
lon_gds attribute 1: standard_name, size = 9 NC_CHAR, value = longitude
lon_gds attribute 2: units, size = 6 NC_CHAR, value = degree
lon_gds attribute 3: purpose, size = 63 NC_CHAR, value = \
 1-D longitude coordinate referred to by geodesic grid variables

In this example lat_gds and lon_gds represent the latitude or longitude, respectively, of cell-based variables. These coordinates (must) have the same single dimension (gds_crd, in this case) as the cell-based variables. And the coordinates must be one-dimensional—multidimensional coordinates will not work.

This infrastructure allows NCO to identify, interpret, and process (e.g., hyperslab) the variables on cell-based grids as easily as it works with regular grids. To time-average all the values between zero and 180 degrees longitude and between plus and minus 30 degress latitude, we use

ncra -O -X 0.,180.,-30.,30. -v gds_3dvar in.nc out.nc

NCO accepts multiple ‘-X’ arguments for cell-based grid multi-slabs, just as it accepts multiple ‘-d’ arguments for multi-slabs of regular coordinates.

ncra -O -X 0.,180.,-30.,30. -X 270.,315.,45.,90. in.nc out.nc

The arguments to ‘-X’ are always interpreted as floating-point numbers, i.e., as coordinate values rather than dimension indices so that these two commands produce identical results

ncra -X 0.,180.,-30.,30. in.nc out.nc
ncra -X 0,180,-30,30 in.nc out.nc

In contrast, arguments to ‘-d’ require decimal places to be recognized as coordinates not indices (see Hyperslabs). We recommend always using decimal points with ‘-X’ arguments to avoid confusion.

3.22 Grid Generation

As of NCO version 4.5.2 (August, 2015), ncks generates accurate and complete SCRIP-format gridfiles for select grid types, including uniform, capped and Gaussian rectangular latitude/longitude grids. The grids are stored in an external grid-file.

All options pertinent to the grid geometry and metadata are passed to NCO via key-value pairs prefixed by the ‘--rgr’ option, or its synonym, ‘--regridding’. The option ‘--rgr’ (and its long option equivalents such as ‘--regridding’) indicates the argument syntax will be key=val. As such, ‘--rgr’ and its synonyms are indicator options that accept arguments supplied one-by-one like ‘--rgr key1=val1 --rgr key2=val2’, or aggregated together in multi-argument format like ‘--rgr key1=val1#key2=val2’ (see Multi-arguments).

The text strings that describe the grid and name the file are important aids to convey the grid geometry to other users. These arguments, and their corresponding keys, are the grid title (grd_ttl), and grid filename (grid), respectively. The numbers of latitudes (lat_nbr) and longitudes (lon_nbr) are independent, and together determine the grid storage size. These four options should be considered mandatory, although NCO provides defaults for any arguments omitted.

The remaining arguments depend on the whether the grid is global or regional. For global grids, one should specify only two more arguments, the latitude (lat_typ) and longitude (lon_typ) grid-types. These types are chosen as described below from a small selection of options that together define the most common rectangular global grids. For regional grids, one must specify the bounding box, i.e., the edges of the rectangular grid on the North (lat_nrt), South (lat_sth), East (lat_est), and West (lat_nrt) sides. Specifying a bounding box for global grids is redundant and will cause an error to ensure the user intends a global grid. NCO assumes that regional grids are uniform, though it will attempt to produce regional grids of other types if the user specifies other latitude (lat_typ) and longitude (lon_typ) grid-types, e.g., Gaussian or Cap. Edges of a regional bounding box may be specified individually, or in the single-argument forms.

The full description of grid-generation arguments, and their corresponding keys, is:

Grid Title: grd_ttl

It is surprisingly difficult to discern the geometric configuration of a grid from the coordinates of a SCRIP-format gridfile. A human-readable grid description should be placed in grd_ttl. Examples include “CAM-FV scalar grid 129x256” and “T42 Gaussian grid”.

Grid File: grid

The grid-generation API was bolted-on to NCO and is somewhat primitive, e.g., having to repeat the ‘--rgr’ option. Another aspect of this is that the output grid filename is distinct from the output filename of the host ncks command. Specify the output gridfile name with grid, e.g., t42_SCRIP.20150901.nc. It is conventional to include a datestamp in the gridfile name. This helps users identify up-to-date and out-of-date grids. Any valid netCDF file may be named as the source (e.g., in.nc). It will not be altered. The destination file (e.g., foo.nc) will be overwritten. Its contents are immaterial.

Grid Types: lat_typ, lon_typ

The keys that hold the longitude and latitude gridtypes (which are, by the way, independent of eachother) are lon_typ and lat_typ. The lat_typ options for global grids are ‘uni’ for Uniform, ‘cap’ (or ‘fv’) for Capped (equivalent to FV), and ‘gss’ for Gaussian. These values are all case-independent, so ‘Gss’ and ‘gss’ both work.

As its name suggests, the latitudes in a Uniform-latitude grid are uniformly spaced ³⁶. The Uniform-latitude grid may have any number of latitudes. NCO can only generate longitude grids (below) that are uniformly spaced, so the Uniform-latitude grids we describe are also uniform in the 2D sense. Uniform grids are intuitive, easy to visualize, and simple to program. Hence their popularity in data exchange, visualization, and archives. Moreover, regional grids (unless they include the poles), are free of polar singularities, and thus are well-suited to storage on Uniform grids. Theoretically, a Uniform-latitude grid could have non-uniform longitudes, but NCO currently does not implement non-uniform longitude grids.

Their mathematical properties (convergence and excessive resolution at the poles, which can appear as singularities) make Uniform grids fraught for use in global models. One purpose Uniform grids serve in modeling is as “offset” or “staggered” grids, meaning grids whose centers are the interfaces of another grid. The Finite-Volume (FV) method is often used to represent and solve the equations of motion in climate-related fields. Many FV solutions (including the popular Lin-Rood method as used in the CESM CAM-FV atmospheric model) evaluate scalar (i.e., non-vector) fields (e.g., temperature, water vapor) at gridcell centers of what is therefore called the scalar grid. FV methods (like Lin-Rood) that employ an Arakawa C-grid or D-grid formulation define velocities on the edges of the scalar grid. This CAM-FV velocity grid is therefore “staggered” or “offset” from the CAM-FV scalar grid by one-half gridcell. The CAM-FV scalar latitude grid has gridpoints (the “caps”) centered on each pole to avoid singularities. The offset of a Cap-grid is a Uniform-grid, so the Uniform grid is often called an FV-”offset” or “staggered” grid. Hence an NCO Uniform grid is equivalent to an NCL “Fixed Offset” grid. For example, a 128x256 Uniform grid is the offset or staggered version of a 129x256 Cap grid (aka FV-grid).

Referring the saucer-like cap-points at the poles, NCO uses the term “Cap grid” to describe the latitude portion of the FV-scalar grid as used by the CAM-FV Lin-Rood dynamics formulation. NCO accepts the shorthand FV, and the more descriptive “Yarmulke”, as synonyms for Cap. A Cap-latitude grid differs from a Uniform-latitude grid in many ways:

Most importantly, Cap grids are 2D-representations of numerical grids with cap-midpoints instead of zonal-teeth convergence at the poles. The rectangular 2D-representation of each cap contains gridcells shaped like sharp teeth that converge at the poles similar to the Uniform grid, but the Cap gridcells are meant to be aggregated into a single cell centered at the pole in a dynamical transport algorithm. In other words, the polar teeth are a convenient way to encode a non-rectangular grid in memory into a rectangular array on disk. Hence Cap grids have the unusual property that the poles are labeled as being both the centers and the outer interfaces of all polar gridcells. Second, Cap grids are uniform in angle except at the poles, where the latitudes span half the meridional range of the rest of the gridcells. Even though in the host dynamical model the Cap grid polar points are melded into caps uniform (in angle) with the rest of the grid, the disk representation on disk is not uniform. Nevertheless, some call the Cap grid a uniform-angle grid because the information contained at the poles is aggregated in memory to span twice the range of a single polar gridcell (which has half the normal width). NCL uses the term “Fixed grid” for a Cap grid. The “Fixed” terminology seems broken.

Finally, Gaussian grids are the Cartesian representation of global spectral transform models. Gaussian grids do not have points at the poles, and must have an even number of latitudes. All three latitude grid-type supported by NCO (Uniform, Cap, and Gaussian) are Regular grids in that they are monotonic.

The lon_typ options for global grids are ‘grn_ctr’ and ‘180_ctr’ for the first gridcell centered at Greenwich or 180 degrees, respecitvely. And ‘grn_wst’ and ‘180_wst’ for Greenwich or 180 degress lying on the western edge of the first gridcell. Many global models use the ‘grn_ctr’ longitude grid as their “scalar grid” (where, e.g., temperature, humidity, and other scalars are defined). The “staggered” or “offset” grid (where often the dynamics variables are defined) then must have the ‘grn_wst’ longitude convention. That way the centers of the scalar grid are the vertices of the offset grid, and visa versa.

Grid Resolution: lat_nbr, lon_nbr

The number of gridcells in the horizontal spatial dimensions are lat_nbr and lon_nbr, respectively. There are no restrictions on lon_nbr for any gridtype. Latitude grids do place some restrictions on lat_nbr (see above). As of NCO version 4.5.3, released in October, 2015, the ‘--rgr latlon=lat_nbr,lon_nbr’ switch may be used to simultaneously specify both latitude and longitude, e.g., ‘--rgr latlon=180,360’.

Grid Edges: lon_wst, lon_est, lat_sth, lat_nrt

The outer edges of a regional rectangular grid are specified by the North (lat_nrt), South (lat_sth), East (lat_est), and West (lat_nrt) sides. Latitudes and longigudes must be specified in degrees (not radians). Latitude edges must be between -90 and 90. Longitude edges may be positive or negative and separated by no more than 360 degrees. The edges may be specified individually with four arguments, consecutively separated by the multi-argument delimiter (‘#’ by default), or together in a short list to the pre-ordered options ‘wesn’ or ‘snwe’. These three specifications are equivalent:

ncks ... --rgr lat_sth=30.0 --rgr lat_nrt=70.0 --rgr lon_wst=-120.0 --rgr lon_est=-90.0 ...
ncks ... --rgr lat_sth=30.0#lat_nrt=70.0#lon_wst=-120.0#lon_est=-90.0 ...
ncks ... --rgr snwe=30.0,70.0,-120.0,-90.0 ...

The first example above supplies the bounding box with four key=val pairs. The second example above supplies the bounding box with a single option in multi-argument format (see Multi-arguments). The third example uses a convenience switch introduced to reduce typing.

Generating common grids:

# 180x360 (1x1 degree) Equi-Angular grid, first longitude centered at Greenwich
ncks --rgr grd_ttl='Equi-Angular grid 180x360' --rgr grid=${DATA}/grids/180x360_SCRIP.20150901.nc \
     --rgr latlon=180,360 --rgr lat_typ=uni --rgr lon_typ=grn_ctr ~zender/nco/data/in.nc ~/foo.nc

# 129x256 CAM-FV grid, first longitude centered at Greenwich
ncks --rgr grd_ttl='CAM-FV scalar grid 129x256' --rgr grid=${DATA}/grids/129x256_SCRIP.20150901.nc \
     --rgr latlon=129,256 --rgr lat_typ=fv --rgr lon_typ=grn_ctr ~zender/nco/data/in.nc ~/foo.nc

# 128x256 Equi-Angular grid, Greenwich west edge of first longitude
# This is the CAM-FV offset grid for the 129x256 CAM-FV scalar grid above
ncks --rgr grd_ttl='Equi-Angular grid 128x256' --rgr grid=${DATA}/grids/128x256_SCRIP.20150901.nc \
     --rgr latlon=128,256 --rgr lat_typ=uni --rgr lon_typ=grn_wst ~zender/nco/data/in.nc ~/foo.nc

# T42 Gaussian grid, first longitude centered at Greenwich
ncks --rgr grd_ttl='T42 Gaussian grid' --rgr grid=${DATA}/grids/t42_SCRIP.20150901.nc \
     --rgr latlon=64,128 --rgr lat_typ=gss --rgr lon_typ=grn_ctr ~zender/nco/data/in.nc ~/foo.nc

# NASA Climate Modeling Grid (CMG) 3600x7200 (0.05x0.05 degree) Equi-Angular grid
# Date-line west edge of first longitude, east edge of last longitude
ncks --rgr grd_ttl='Equi-Angular grid 3600x7200 (NASA CMG)' --rgr grid=${DATA}/grids/3600x7200_SCRIP.20160301.nc \
     --rgr latlon=3600,7200 --rgr lat_typ=uni --rgr lon_typ=180_wst ~zender/nco/data/in.nc ~/foo.nc

Often researchers face the problem not of generating a known, idealized grid but of understanding an unknown, possibly irregular or curvilinear grid underlying a dataset produced elsewhere. NCO will infer the grid of a datafile by examining its coordinates (and boundaries, if available), reformat that information as necessary to diagnose gridcell areas, and output the results in SCRIP format. As of NCO version 4.5.3, released in October, 2015, the ‘--rgr nfr='y'’ switch activates the machinery to infer the grid rather than construct the grid from other user-specified switches. To infer the grid properties, NCO interrogates input-file for horizontal coordinate information, such as the presence of dimension names rooted in latitude/longitude-naming traditions and conventions. Once NCO identifies the likely horizontal dimensions it looks for horizontal coordinates and bounds. If bounds are not found, NCO assumes the underlying grid comprises quadrilateral cells whose edges are midway between cell centers, for both rectilinear and curvilinear grids.

# Infer the grid of the AIRS swath from input, write it to grd_airs.nc
ncks --rgr nfr=y --rgr grid=${DATA}/sld/rgr/grd_airs.nc \
     ${DATA}/sld/raw/AIRS.2014.10.01.202.L2.TSurfStd.Regrid010.1DLatLon.nc ~/foo.nc

When inferring grids, the grid file (grd_airs.nc) is written, the input file (AIRS...nc) is read, and the output file (foo.nc) is overwritten (its contents are immaterial).

Another task that arises in regridding is characterizing new grids. In such cases it can be helpful to have a “skeleton” version of a dataset on the grid, so that grid center and interfaces locations can be assessed, continental outlines can be examined, or the skeleton can be manually populated with data rather than relying on a model. SCRIP files can be difficult to visualize and manipulate, so NCO will provide, if requested, a so-called skeleton file on the user-specified grid. As of NCO version 4.5.3, released in October, 2015, the ‘--rgr skl=fl_skl’ switch outputs the skeleton file to fl_skl. The skeleton file may then be examined in a dataset viewer, populated with data, and generally serve as a template for what to expect from datasets of the same geometry.

# Generate T42 Gaussian grid file t42_SCRIP.nc and skeleton file t42_skl.nc
ncks --rgr skl=${DATA}/grids/t42_skl.nc --rgr grid=${DATA}/grids/t42_SCRIP.nc \
     --rgr latlon=64,128 --rgr lat_typ=gss --rgr lon_typ=Grn_ctr \
     ~zender/nco/data/in.nc ~/foo.nc

When generating skeleton files, both the grid file (t42_SCRIP.nc) and the skeleton file (t42_skl.nc) are written, the input file (in.nc) is ignored, and the output file (foo.nc) is overwritten (its contents are immaterial).

3.23 Regridding

NCO includes extensive regridding features in ncclimo (as of version 4.6.0 in May, 2016), ncremap (as of version 4.5.4 in November, 2015) and ncks (since version 4.5.0 in June, 2015). Regridding can involve many choices, options, inputs, and outputs. The appropriate operator for this workflow is the ncremap script which automatically handles many details of regridding and passes the required commands to ncks and external programs. Occasionally users need access to lower-level remapping functionality present in ncks and not exposed to direct manipulation through ncremap or ncclimo. This section describes the lower-level functionality and switches as implemented in ncks. Knowing what these features are will help ncremap and ncclimo users understand the full potential of these operators.

ncks supports horizontal regridding of datasets where the grids and weights are all stored in an external map-file. Use the ‘--map’ or ‘--rgr_map’ options to specify the map-file, and NCO will regrid the input-file to a new (or possibly the same, aka, an identity mapping) horizontal grid in the output-file, using the input and output grids and mapping weights specified in the ESMF- or SCRIP-format map-file. Currently NCO understands only the mapfile format pioneered by SCRIP (http://oceans11.lanl.gov/svn/SCRIP/trunk/SCRIP) and later extended by ESMF (http://www.earthsystemcog.org/projects/regridweightgen), and adopted (along with Exodus) by TempestRemap (https://github.com/ClimateGlobalChange/tempestremap.git). See those references for documentation on map formats, grid specification, and weight generation.

The regridding currently supported by NCO could equally well be called weight-application. NCO reads-in pre-stored weights from the map-file and applies them to (almost) every variable, thereby creating a regridded output-file. Specify regridding with a standard ncks command and options along with the additional specification of a map-file:

# Regrid entire file, same output format as input:
ncks --map=map.nc in.nc out.nc
# Entire file, netCDF4 output:
ncks -4 --map=map.nc in.nc out.nc
# Deflated netCDF4 output
ncks -4 -L 1 --map=map.nc in.nc out.nc
# Selected variables
ncks -v FS.?,T --map=map.nc in.nc out.nc
# Threading
ncks -t 8 --map=map.nc in.nc out.nc
# Deflated netCDF4 output, threading, selected variables:
ncks -4 -L 1 -t 8 -v FS.?,T --map=map.nc in.nc out.nc

OpenMP threading works well with regridding large datasets. Threading improves throughput of regridding 1–10 GB files by factors of 2–5. Options specific to regridding are described below.

NCO supports 1D⇒1D, 1D⇒2D, 2D⇒1D, and 2D⇒2D regridding for any unstructured 1D-grid and any rectangular 2D-grid. This has been tested by converting among and between Gaussian, equiangular, FV, unstructured cubed-sphere grids, and regionally refined grids. Support for irregular 2D- and regional grids (e.g., swath-like data) is planned.

Conservative regridding is, for first-order accurate algorithms, a straightforward procedure of identifying gridcell overlap and apportioning values correctly from source to destination. The presence of missing values forces a decision on how to handle destination gridcells where some but not all source cells are valid. NCO allows the user to choose between two distinct algorithms: “conservative” and “renormalized”. The “conservative” algorithm uses all valid data from the input grid on the output grid once and only once. Destination cells receive the weighted valid values of the source cells. This is conservative because the global integrals of the source and destination fields are equal. The “renormalized” algorithm divides the destination value by the sum of the valid weights. This produces values equal to the mean of the valid input values, but extended to the entire destination gridcell. Thus renormalization is equivalent to extrapolating valid data to missing regions. Input and output integrals are unequal and renormalized regridding is not conservative. Both algorithms produce identical answers when no missing data maps to the destination gridcell.

By default, NCO implements the “conservative” algorithm because it has useful properties, is simpler to understand, and requires no additional parameters. To employ the “renormalized” algorithm instead, use the ‘--rnr’, ‘--rgr_rnr’, or ‘--renormalize’ options to supply wgt_thr, the threshold weight for valid destination values. Valid values must cover at least the fraction wgt_thr of the destination gridcell to meet the threshold for a non-missing destination value. When wgt_thr is exceeded, the mean valid value is renormalized by the valid area and placed in the destination gridcell. If the valid area covers less than wgt_thr, then the destination gridcell is assigned the missing value. Valid values of wgt_thr range from zero to one. Keep in mind though, that this threshold is potentially a divisor, and values of zero or very near to zero can lead to floating-point underflow and divide-by-zero errors. For convenience NCO permits users to specify a wgt_thr = 0.0 threshold weight. This indicates that any valid data should be represented and renormalized on the output grid.

ncks           --map=map.nc in.nc out.nc # Conservative regridding
ncks --rnr=0.1 --map=map.nc in.nc out.nc # Renormalized regridding

The first example uses the default conservative algorithm. The second example specifies that valid values must cover at least 10% of the destination gridcell to meet the threshold for a non-missing destination value. With valid destination areas of, say 25% or 50%, the renormalized algorithm would produce destination values greater than the conservative algorithm by factors of four or two, respectively.

In practice, it may make sense to use the default “conservative” algorithm when performing conservative regridding, and the “renormalized” algorithm when performing other regridding such as bilinear interpolation or nearest-neighbor. Another consideration is whether the fields being regridded are fluxes or state variables. For example, temperature (unlike heat) or concentrations (amount per unit volume) are not physically conserved quantities under areal-regridding so it often makes sense to interpolate them in a non-conservative fashion, to preserve their fine-scale structure. Few researchers can digest the unphysical values of temperature that the “conservative” option will produce in regions rife with missing values. A counter-example is fluxes, which should be physically conserved under areal-regridding. One should consider both the type of field and its conservation properties before choosing a regridding strategy.

NCO automatically annotates the output with relevant metadata such as coordinate bounds, axes, and vertices (à la CF). These annotations include

Horizontal Dimension Names: lat, lon

The name of the horizontal spatial dimensions assumed to represent latitude and longitude in 2D rectangular input files are lat_nm and lon_nm, which default to lat and lon, respectively. Variables that contain a lat_nm-dimension and a lon_nm-dimension on a 2D-rectangular input grid will be regridded, and variables regridded to a 2D-rectangular output grid will all contain the lat_nm- and lon_nm-dimensions, and variables regridded to a 1D-unstructured output grid will have lat_nm and lon_nm as auxiliary coordinate variables. To treat different dimensions and variables as latitude and longitude, use the options ‘--rgr lat_nm=lat_nm’ and ‘--rgr lon_nm=lon_nm’. Note that, for now at least, lat_nm and lon_nm indicate both the variable names associated and, where applicable (i.e., on 2D-grids), the dimensions of the horizontal coordinates.

Unstructured Dimension Name: col

The name of the horizontal spatial dimension assumed to delineate an unstructured grid is col_nm, which defaults to ncol (number of columns), the name CAM employs. Other common names for the columns in an unstructured grid include lndgrid (used by CLM), and nCells (used by MPAS-O). Variables that contain the col_nm-dimension on an unstructured input grid will be regridded, and regridded variables written to an unstructured output grid will all contain the col_nm-dimension. To treat a different dimension as unstructured, use the option ‘--rgr col_nm=col_nm’. Note: Often there is no coordinate variable for the col_nm-dimension, i.e., there is no variable named col_nm, although such a coordinate could contain useful information about the unstructured grid.

Structured Grid Standard Names and Units

Longitude and latitude coordinates (both regular and auxiliary, i.e., for unstructured grids) receive CF standard_name values of latitude and longitude, CF axes attributes with values X and Y, and units attributes with values degrees_east and degrees_north, respectively.

Unstructured Grid Auxiliary Coordinates

Unstructured grid auxiliary coordinates for longitude and latitude receive CF coordinates attributes with values lon and lat, respectively.

Structured Grid Bounds Variables: bnd, lat_bnd, lon_bnd

Structured grids with 1D-coordinates use the dimension bnd_nm (which defaults to nbnd) with the spatial bounds variables in lat_bnd_nm and lon_bnd_nm which default to lon_bnds and lat_bnds, respectively. By default spatial bounds for such structured grids parallel the oft-used temporal bounds dimension (nbnd=2) and variable (time_bnds). Bounds are attached to the horizontal spatial dimensions via their bounds attributes. Change the spatial bounds dimension with the option ‘--rgr bnd_nm=bnd_nm’. Rename the spatial bounds variables with the options ‘--rgr lat_bnd_nm=lat_bnd_nm’ and ‘--rgr lon_bnd_nm=lon_bnd_nm’.

Unstructured Grid Bounds Variables: bnd, lat_bnd, lon_bnd

Unstructured grids with 1D-coordinates use the dimension bnd_nm (which defaults to nv, number of vertices) for the spatial bounds variables lat_bnd_nm and lon_bnd_nm which default to lat_vertices and lon_vertices, respectively. It may be impossible to re-use the temporal bounds dimension (often nbnd) for unstructure grids, because the gridcells are not rectangles, and thus require specification of all vertices for each gridpoint, rather than only two parallel interfaces per dimension. These bounds are attached to the horizontal spatial dimensions via their bounds attributes. Change the spatial bounds dimension with the option ‘--rgr bnd_nm=bnd_nm’. Rename the spatial bounds variables with the options ‘--rgr lat_bnd_nm=lat_bnd_nm’ and ‘--rgr lon_bnd_nm=lon_bnd_nm’. The temporal bounds dimension in unstructured grid output remains as in the input-file, usually nbnd.

Gridcell Area: area

The variable area_nm (which defaults to area) is always (re-)created in the output_file to hold the gridcell area in steradians. To store the area in a different variable, use the option ‘--rgr area=area_nm’. The area_nm variable receives a standard_name attribute of cell_area, a units attribute of steradian (the SI unit of solid angle), and a cell_methods attribute with value lat, lon: sum, which indicates that area_nm is extensive, meaning that its value depends on the gridcell boundaries. Since area_nm is a property of the grid, it is read directly from the map-file rather than regridded itself.

Gridcell Fraction: frc

The variable frc_nm (which defaults to frac_b) is automatically copied to the output_file to hold the valid fraction of each gridcell when certain conditions are met. First, the regridding method must be conservative. Second, at least one value of frc_nm must be non-unity. These conditions ensure that whenever fractional gridcells affect the regridding, they are also placed in the output file. To store the fraction in a different variable, use the option ‘--rgr frc_nm=frc_nm’. The frc_nm variable receives a cell_methods attribute with value lat, lon: sum, which indicates that frc_nm is extensive, meaning that its value depends on the gridcell boundaries. Since frc_nm is a property of the grid, it is read directly from the map-file rather than regridded itself.

Latitude weights: lat_wgt

Rectangular 2D-grids use the variable lat_wgt_nm, which defaults to gw (originally for “Gaussian weight”), to store the 1D-weight appropriate for area-weighting the latitude grid. To store the latitude weight in a different variable, use the option ‘--rgr lat_wgt=lat_wgt_nm’. The lat_wgt_nm variable will not appear in 1D-grid output. Weighting statistics by latitude (i.e., by lat_wgt_nm will produce the same answers (up-to round-off error) as weighting by area (i.e., by area_nm) in grids that have both variables. The former requires less memory because lat_wgt_nm is 1D), whereas the latter is more general because area_nm works on any grid.

Provenance Attributes

The map-file and input-file names are stored in the output-file global attributes mapping_file and source_file, respectively.

One may supply muliple ‘--rgr key=value’ options to simultaneously customize multiple grid-field names. The following examples may all be assumed to end with the standard options ‘--map=map.nc in.nc out.nc’.

ncks --rgr lat_nm=latitude --rgr lon_nm=longitude
ncks --rgr col_nm=column --rgr lat_wgt=lat_wgt
ncks --rgr bnd_nm=bounds --rgr lat_bnd_nm=lat_bounds --rgr lon_bnd_nm=lon_bounds
ncks --rgr bnd_nm=vertices --rgr lat_bnd_nm=lat_vrt --rgr lon_bnd_nm=lon_vrt

The first command causes the regridder to associate the latitude and longitude dimensions with the dimension names latitude and longitude (instead of the defaults, lat and lon). The second command causes the regridder to associate the independent columns in an unstructured grid with the dimension name column (instead of the default, ncol) and the variable containing latitude weights to be named lat_wgt (instead of the default, gw). The third command associates the latitude and longitude bounds with the dimension bounds (instead of the default, nbnd) and the variables lat_bounds and lon_bounds (instead of the defaults, lat_bnds and lon_bnds, respectively). The fourth command associates the latitude and longitude bounds with the dimension vertices (instead of the default, nv) and the variables lat_vrt and lon_vrt (instead of the defaults, lat_vertices and lon_vertices, respectively).

When used with an identity remapping files, regridding can signficantly enhance the metadata and therefore the dataset usability. Consider these selected metadata (those unchanged are not shown for brevity) associated with the variable FSNT from typical unstructured grid (CAM-SE cubed-sphere) output before and after an identity regridding:

# Raw model output before regridding
netcdf ne30_FSNT {
  dimensions:
    nbnd = 2 ;
    ncol = 48602 ;
    time = UNLIMITED ; // (1 currently)

  variables:
    float FSNT(time,ncol) ;
      FSNT:long_name = "Net solar flux at top of model" ;

    double time(time) ;
      time:long_name = "time" ;
      time:bounds = "time_bnds" ;

    double time_bnds(time,nbnd) ;
      time_bnds:long_name = "time interval endpoints" ;
} // group /

# Same model output after identity regridding
netcdf dogfood {
  dimensions:
    nbnd = 2 ;
    ncol = 48602 ;
    nv = 5 ;
    time = 1 ;

  variables:
    float FSNT(time,ncol) ;
      FSNT:long_name = "Net solar flux at top of model" ;
      FSNT:coordinates = "lat lon" ;

    double lat(ncol) ;
      lat:long_name = "latitude" ;
      lat:standard_name = "latitude" ;
      lat:units = "degrees_north" ;
      lat:axis = "Y" ;
      lat:bounds = "lat_vertices" ;
      lat:coordinates = "lat lon" ;

    double lat_vertices(ncol,nv) ;
      lat_vertices:long_name = "gridcell latitude vertices" ;

    double lon(ncol) ;
      lon:long_name = "longitude" ;
      lon:standard_name = "longitude" ;
      lon:units = "degrees_east" ;
      lon:axis = "X" ;
      lon:bounds = "lon_vertices" ;
      lon:coordinates = "lat lon" ;

    double lon_vertices(ncol,nv) ;
      lon_vertices:long_name = "gridcell longitude vertices" ;

    double time(time) ;
      time:long_name = "time" ;
      time:bounds = "time_bnds" ;

    double time_bnds(time,nbnd) ;
      time_bnds:long_name = "time interval endpoints" ;
} // group /

The raw model output lacks the CF coordinates and bounds attributes that the regridder adds. The metadata turns lat and lon into auxiliary coordinate variables (see Auxiliary Coordinates) which can then be hyperslabbed (with ‘-X’) using latitude/longitude coordinates bounding the region of interest:

% ncks -u -H -X 314.6,315.3,-35.6,-35.1 -v FSNT dogfood.nc
time[0]=31 ncol[0] FSNT[0]=344.575 W/m2

ncol[0] lat[0]=-35.2643896828 degrees_north

ncol[0] nv[0] lat_vertices[0]=-35.5977213708 
ncol[0] nv[1] lat_vertices[1]=-35.5977213708 
ncol[0] nv[2] lat_vertices[2]=-35.0972113817 
ncol[0] nv[3] lat_vertices[3]=-35.0972113817 
ncol[0] nv[4] lat_vertices[4]=-35.0972113817 

ncol[0] lon[0]=315 degrees_east

ncol[0] nv[0] lon_vertices[0]=315 
ncol[0] nv[1] lon_vertices[1]=315 
ncol[0] nv[2] lon_vertices[2]=315.352825437 
ncol[0] nv[3] lon_vertices[3]=314.647174563 
ncol[0] nv[4] lon_vertices[4]=314.647174563 

time[0]=31 days since 1979-01-01 00:00:00

time[0]=31 nbnd[0] time_bnds[0]=0 
time[0]=31 nbnd[1] time_bnds[1]=31 

Thus auxiliary coordinate variables help to structure unstructured grids. The expanded metadata annotations from an identity regridding may obviate the need to place unstructured data on a rectangular grid. For example, statistics for regions that can be expressed as unions of rectangular regions can now be performed on the native (unstructured) grid.

Here are some quick examples of regridding from common models. All examples require ‘in.nc out.nc’ at the end.

# Identity re-map ACME CAM-SE Cubed-Sphere output (to improve metadata)
ncks --map=${DATA}/maps/map_ne30np4_to_ne30np4_aave.20150603.nc
# Convert ACME CAM-SE Cubed Sphere output to rectangular lat/lon
ncks --map=${DATA}/maps/map_ne30np4_to_fv129x256_aave.150418.nc
# Convert CAM3 T42 output to Cubed-Sphere grid
ncks --map=${DATA}/maps/map_ne30np4_to_t42_aave.20150601.nc

3.24 UDUnits Support

There is more than one way to hyperskin a cat. The UDUnits package provides a library which, if present, NCO uses to translate user-specified physical dimensions into the physical dimensions of data stored in netCDF files. Unidata provides UDUnits under the same terms as netCDF, so sites should install both. Compiling NCO with UDUnits support is currently optional but may become required in a future version of NCO.

Two examples suffice to demonstrate the power and convenience of UDUnits support. First, consider extraction of a variable containing non-record coordinates with physical dimensions stored in MKS units. In the following example, the user extracts all wavelengths in the visible portion of the spectrum in terms of the units very frequently used in visible spectroscopy, microns:

% ncks -C -H -v wvl -d wvl,"0.4 micron","0.7 micron" in.nc
wvl[0]=5e-07 meter

The hyperslab returns the correct values because the wvl variable is stored on disk with a length dimension that UDUnits recognizes in the units attribute. The automagical algorithm that implements this functionality is worth describing since understanding it helps one avoid some potential pitfalls. First, the user includes the physical units of the hyperslab dimensions she supplies, separated by a simple space from the numerical values of the hyperslab limits. She encloses each coordinate specifications in quotes so that the shell does not break the value-space-unit string into separate arguments before passing them to NCO. Double quotes ("foo") or single quotes ('foo') are equally valid for this purpose. Second, NCO recognizes that units translation is requested because each hyperslab argument contains text characters and non-initial spaces. Third, NCO determines whether the wvl is dimensioned with a coordinate variable that has a units attribute. In this case, wvl itself is a coordinate variable. The value of its units attribute is meter. Thus wvl passes this test so UDUnits conversion is attempted. If the coordinate associated with the variable does not contain a units attribute, then NCO aborts. Fourth, NCO passes the specified and desired dimension strings (microns are specified by the user, meters are required by NCO) to the UDUnits library. Fifth, the UDUnits library that these dimension are commensurate and it returns the appropriate linear scaling factors to convert from microns to meters to NCO. If the units are incommensurate (i.e., not expressible in the same fundamental MKS units), or are not listed in the UDUnits database, then NCO aborts since it cannot determine the user’s intent. Finally, NCO uses the scaling information to convert the user-specified hyperslab limits into the same physical dimensions as those of the corresponding cooridinate variable on disk. At this point, NCO can perform a coordinate hyperslab using the same algorithm as if the user had specified the hyperslab without requesting units conversion.

The translation and dimensional innterpretation of time coordinates shows a more powerful, and probably more common, UDUnits application. In this example, the user prints all data between 4 PM and 7 PM on December 8, 1999, from a variable whose time dimension is hours since the year 1900:

% ncks -u -H -C -v time_udunits -d time_udunits,"1999-12-08 \
  16:00:0.0","1999-12-08 19:00:0.0" in.nc
time_udunits[1]=876018 hours since 1900-01-01 00:00:0.0

Here, the user invokes the stride (see Stride) capability to obtain every other timeslice. This is possible because the UDUnits feature is additive, not exclusive—it works in conjunction with all other hyperslabbing (see Hyperslabs) options and in all operators which support hyperslabbing. The following example shows how one might average data in a time period spread across multiple input files

ncra -d time,"1939-09-09 12:00:0.0","1945-05-08 00:00:0.0" \
  in1.nc in2.nc in3.nc out.nc

Note that there is no excess whitespace before or after the individual elements of the ‘-d’ argument. This is important since, as far as the shell knows, ‘-d’ takes only one command-line argument. Parsing this argument into its component dim,[min][,[max][,[stride]]] elements (see Hyperslabs) is the job of NCO. When unquoted whitespace is present between these elements, the shell passes NCO arugment fragments which will not parse as intended.

NCO implemented support for the UDUnits2 library with version 3.9.2 (August, 2007). The UDUnits2 package supports non-ASCII characters and logarithmic units. We are interested in user-feedback on these features.

One aspect that deserves mention is that UDUnits, and thus NCO, supports run-time definition of the location of the relevant UDUnits databases. With UDUnits version 1, users may specify the directory which contains the UDUnits database, udunits.dat, via the UDUNITS_PATH environment variable. With UDUnits version 2, users may specify the UDUnits database file itself, udunits2.xml, via the UDUNITS2_XML_PATH environment variable.

# UDUnits1
export UDUNITS_PATH='/unusual/location/share/udunits'
# UDUnits2
export UDUNITS2_XML_PATH='/unusual/location/share/udunits/udunits2.xml'

This run-time flexibility can enable the full functionality of pre-built binaries on machines with libraries in different locations.

The UDUnits package documentation describes the supported formats of time dimensions. Among the metadata conventions that adhere to these formats are the Climate and Forecast (CF) Conventions and the Cooperative Ocean/Atmosphere Research Data Service (COARDS) Conventions. The following ‘-d arguments’ extract the same data using commonly encountered time dimension formats:

-d time,'1918-11-11 00:00:0.0','1939-09-09 00:00:0.0'
-d time,'1918-11-11 00:00:0.0','1939-09-09 00:00:0.0'
-d time,'1918-11-11T00:00:0.0Z','1939-09-09T00:00:0.0Z'
-d time,'1918-11-11','1939-09-09'
-d time,'1918-11-11','1939-9-9'

All of these formats include at least one dash - in a non-leading character position (a dash in a leading character position is a negative sign). NCO assumes that a space, colon, or non-leading dash in a limit string indicates that a UDUnits units conversion is requested. Some date formats like YYYYMMDD that are valid in UDUnits are ambiguous to NCO because it cannot distinguish a purely numerical date (i.e., no dashes or text characters in it) from a coordinate or index value:

-d time,1918-11-11 # Interpreted as the date November 11, 1918
-d time,19181111   # Interpreted as time-dimension index 19181111
-d time,19181111.  # Interpreted as time-coordinate value 19181111.0

Hence, use the YYYY-MM-DD format rather than YYYYMMDD for dates.

As of version 4.0.0 (January, 2010), NCO supports some calendar attributes specified by the CF conventions.

Supported types:

"365_day"/"noleap", "360_day", "gregorian", "standard"

Unsupported types:

"366_day"/"all_leap","proleptic_gregorian","julian","none"

Unsupported types default to mixed Gregorian/Julian as defined by UDUnits.

An Example: Consider the following netCDF variable

variables:
  double lon_cal(lon_cal) ;
    lon_cal:long_name = "lon_cal" ;
    lon_cal:units = "days since 1964-2-28 0:0:0" ;
    lon_cal:calendar = "365_day" ;
data:
  lon_cal = 1,2,3,4,5,6,7,8,9,10;

‘ncks -v lon_cal -d lon_cal,'1964-3-1 0:00:0.0','1964-3-4 00:00:0.0'’ results in lon_cal=1,2,3,4.

netCDF variables should always be stored with MKS (i.e., God’s) units, so that application programs may assume MKS dimensions apply to all input variables. The UDUnits feature is intended to alleviate some of the NCO user’s pain when handling MKS units. It connects users who think in human-friendly units (e.g., miles, millibars, days) to extract data which are always stored in God’s units, MKS (e.g., meters, Pascals, seconds). The feature is not intended to encourage writers to store data in esoteric units (e.g., furlongs, pounds per square inch, fortnights).

3.25 Rebasing Time Coordinate

Time rebasing is invoked when numerous files share a common record coordinate, and the record coordinate units change among input files. The rebasing is performed automatically if and only if UDUnits is installed. Usually rebasing occurs when the recoordinate is a time-based variable, and times are recorded in units of a time-since-basetime, and the basetime changes from file to file. Since the output file can have only one unit (i.e., one basetime) for the record coordinate, NCO, in such cases, chooses the units of the first input file to be the units of the output file. It is necessary to “rebase” all the input record variables to this output time unit in order for the output file to have the correct values.

For example suppose the time coordinate is in hours and each day in January is stored in its own daily file. Each daily file records the temperature variable tpt(time) with an (unadjusted) time coordinate value between 0–23 hours, and uses the units attribute to advance the base time:

file01.nc time:units="hours since 1990-1-1"   
file02.nc time:units="hours since 1990-1-2"   
...
file31.nc time:units="hours since 1990-1-31"   

// Mean noontime temperature in January
ncra -v tpt -d time,"1990-1-1 12:00:00","1990-1-31 23:59:59",24 \
      file??.nc noon.nc    

// Concatenate day2 noon through day3 noon records
ncrcat -v tpt -d time,"1990-1-2 12:00:00","1990-1-3 11:59:59" \ 
      file01.nc file02.nc file03.nc noon.nc    

// Results: time is "re-based" to the time units in "file01.nc"
time=36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, \
     51, 52, 53, 54, 55, 56, 57, 58, 59 ;
  
// If we repeat the above command but with only two input files...
ncrcat -v tpt -d time,"1990-1-2 12:00:00","1990-1-3 11:59:59" \
      file02.nc file03 noon.nc    

// ...then output time coordinate is based on time units in "file02.nc"
time = 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, \ 
     26, 27, 28, 29, 30, 31, 32, 33, 34, 35 ;

As of NCO version 4.2.1 (August, 2012), NCO automatically rebases not only the record coordinate (time, here) but also any cell boundaries associated with the record coordinate (e.g., time_bnds) (see CF Conventions). As of NCO version 4.4.9 (May, 2015), NCO also rebases any climatology boundaries associated with the record coordinate (e.g., climatology_bounds) (see CF Conventions).

3.26 Multiple Record Dimensions

The netCDF3 file format allows only one record dimension, and that dimension must be the first dimension (i.e., the least rapidly varying dimension) of any variable in which it appears. This imposes certain rules on how operators must perform operations that alter the ordering of dimensions or the number of record variables. The netCDF4 file format has no such restrictions. Files and variables may have any number of record dimensions in any order. This additional flexibility of netCDF4 can only be realized by selectively abandoning the constraints that would make operations behave completely consistently between netCDF3 and netCDF4 files.

NCO chooses, by default, to impose netCDF3-based constraints on netCDF4 files. This reduces the number of unanticipated consequences and keeps the operators functioning in a familiar way. Put another way, NCO limits production of additional record dimensions so processing netCDF4 files leads to the same results as processing netCDF3 files. Users can override this default with the ‘--mrd’ (or ‘--multiple_record_dimension’) switch, which enables netCDF4 variables to accumulate additional record dimensions.

How can additional record dimensions be produced? Most commonly ncecat (in record-aggregate mode) defines a new leading record dimension. In netCDF4 files this becomes an additional record dimension unless the original record dimension is changed to a fixed dimension (as must be done in netCDF3 files). Also when ncpdq reorders dimensions it can preserve the “record” property of record variables. ncpdq tries to define as a record dimension whichever dimension ends up first in a record variable, and, in netCDF4 files, this becomes an additional record dimension unless the original record dimension is changed to a fixed dimension (as must be done in netCDF3 files). It it easier if ncpdq and ncecat do not increase the number of record dimensions in a variable so that is the default. Use ‘--mrd’ to override this.

3.27 Missing values

The phrase missing data refers to data points that are missing, invalid, or for any reason not intended to be arithmetically processed in the same fashion as valid data. The NCO arithmetic operators attempt to handle missing data in an intelligent fashion. There are four steps in the NCO treatment of missing data:

1. Identifying variables that may contain missing data.

   NCO follows the convention that missing data should be stored with the _FillValue specified in the variable’s _FillValue attributes. The only way NCO recognizes that a variable may contain missing data is if the variable has a _FillValue attribute. In this case, any elements of the variable which are numerically equal to the _FillValue are treated as missing data.

   NCO adopted the behavior that the default attribute name, if any, assumed to specify the value of data to ignore is _FillValue with version 3.9.2 (August, 2007). Prior to that, the missing_value attribute, if any, was assumed to specify the value of data to ignore. Supporting both of these attributes simultaneously is not practical. Hence the behavior NCO once applied to missing_value it now applies to any _FillValue. NCO now treats any missing_value as normal data ³⁷.

   It has been and remains most advisable to create both _FillValue and missing_value attributes with identical values in datasets. Many legacy datasets contain only missing_value attributes. NCO can help migrating datasets between these conventions. One may use ncrename (see ncrename netCDF Renamer) to rename all missing_value attributes to _FillValue:

   ncrename -a .missing_value,_FillValue inout.nc
   

   Alternatively, one may use ncatted (see ncatted netCDF Attribute Editor) to add a _FillValue attribute to all variables

   ncatted -O -a _FillValue,,o,f,1.0e36 inout.nc
   

2. Converting the _FillValue to the type of the variable, if neccessary.

   Consider a variable var of type var_type with a _FillValue attribute of type att_type containing the value _FillValue. As a guideline, the type of the _FillValue attribute should be the same as the type of the variable it is attached to. If var_type equals att_type then NCO straightforwardly compares each value of var to _FillValue to determine which elements of var are to be treated as missing data. If not, then NCO converts _FillValue from att_type to var_type by using the implicit conversion rules of C, or, if att_type is NC_CHAR ³⁸, by typecasting the results of the C function strtod(_FillValue). You may use the NCO operator ncatted to change the _FillValue attribute and all data whose data is _FillValue to a new value (see ncatted netCDF Attribute Editor).

3. Identifying missing data during arithmetic operations.

   When an NCO arithmetic operator processes a variable var with a _FillValue attribute, it compares each value of var to _FillValue before performing an operation. Note the _FillValue comparison imposes a performance penalty on the operator. Arithmetic processing of variables which contain the _FillValue attribute always incurs this penalty, even when none of the data are missing. Conversely, arithmetic processing of variables which do not contain the _FillValue attribute never incurs this penalty. In other words, do not attach a _FillValue attribute to a variable which does not contain missing data. This exhortation can usually be obeyed for model generated data, but it may be harder to know in advance whether all observational data will be valid or not.

4. Treatment of any data identified as missing in arithmetic operators.

   NCO averagers (ncra, nces, ncwa) do not count any element with the value _FillValue towards the average. ncbo and ncflint define a _FillValue result when either of the input values is a _FillValue. Sometimes the _FillValue may change from file to file in a multi-file operator, e.g., ncra. NCO is written to account for this (it always compares a variable to the _FillValue assigned to that variable in the current file). Suffice it to say that, in all known cases, NCO does “the right thing”.

   It is impossible to determine and store the correct result of a binary operation in a single variable. One such corner case occurs when both operands have differing _FillValue attributes, i.e., attributes with different numerical values. Since the output (result) of the operation can only have one _FillValue, some information may be lost. In this case, NCO always defines the output variable to have the same _FillValue as the first input variable. Prior to performing the arithmetic operation, all values of the second operand equal to the second _FillValue are replaced with the first _FillValue. Then the arithmetic operation proceeds as normal, comparing each element of each operand to a single _FillValue. Comparing each element to two distinct _FillValue’s would be much slower and would be no likelier to yield a more satisfactory answer. In practice, judicious choice of _FillValue values prevents any important information from being lost.

3.28 Chunking

All netCDF4-enabled NCO operators that define variables support a plethora of chunksize options. Chunking can significantly accelerate or degrade read/write access to large datasets. Dataset chunking issues are described by THG and Unidata here, here, and here. NCO authors are working on generalized algorithms and applications of chunking strategies (stay tuned for more in 2017).

The NCO chunking implementation is designed to be flexible. Users control four aspects of the chunking implementation. These are the chunking policy, chunking map, chunksize, and minimum chunksize. The chunking policy determines which variables to chunk, and the chunking map determines how (with what exact sizes) to chunk those variables. These are high-level mechanisms that apply to an entire file and all variables and dimensions. The chunksize option allows per-dimension specification of sizes that will override the selected (or default) chunking map.

The distinction between elements and bytes is subtle yet crucial to understand. Elements refers to values of an array, whereas bytes refers to the memory size required to hold the elements. These measures differ by a factor of four or eight for NC_FLOAT or NC_DOUBLE, respectively. The option ‘--cnk_scl’ takes an argument sz_lmn measured in elements. The options ‘--cnk_byt’ and ‘--cnk_min’ take arguments sz_byt measured in bytes.

Use the ‘--cnk_min=sz_byt’ option to set the minimum size in bytes (not elements) of variables to chunk. This threshold is intended to restrict use of chunking to variables for which it is efficient. By default this minimum variable size for chunking is twice the system blocksize (when available) and is 8192 bytes otherwise. Users may set this to any value with the ‘--cnk_min=sz_byt’ switch. To guarantee that chunking is performed on all arrays, regardless of size, set the minimum size to one byte (not to zero bytes).

The chunking implementation is similar to a hybrid of the ncpdq packing policies (see ncpdq netCDF Permute Dimensions Quickly) and hyperslab specifications (see Hyperslabs). Each aspect is intended to have a sensible default, so that many users only need to set one switch to obtain sensible chunking. Power users can tune chunking with the three switches in tandem to obtain optimal performance.

By default, NCO preserves the chunking characteristics of the input file in the output file ³⁹. In other words, preserving chunking requires no switches or user intervention.

Users specify the desired chunking policy with the ‘-P’ switch (or its long option equivalents, ‘--cnk_plc’ and ‘--chunk_policy’) and its cnk_plc argument. As of August, 2014, six chunking policies are implemented:

Chunk All Variables

Definition: Chunk all variables possible. For obvious reasons, scalar variables cannot be chunked.
Alternate invocation: ncchunk
cnk_plc key values: ‘all’, ‘cnk_all’, ‘plc_all’
Mnemonic: All

Chunk Variables with at least Two Dimensions [default]

Definition: Chunk all variables possible with at least two dimensions
Alternate invocation: none
cnk_plc key values: ‘g2d’, ‘cnk_g2d’, ‘plc_g2d’
Mnemonic: Greater than or equal to 2 Dimensions

Chunk Variables with at least Three Dimensions

Definition: Chunk all variables possible with at least three dimensions
Alternate invocation: none
cnk_plc key values: ‘g3d’, ‘cnk_g3d’, ‘plc_g3d’
Mnemonic: Greater than or equal to 3 Dimensions

Chunk One-Dimensional Record Variables

Definition: Chunk all 1-D record variables
Alternate invocation: none
Any specified (with ‘--cnk_dmn’) record dimension chunksizes will be applied only to 1-D record variables (and to no other variables). Other dimensions may be chunked with their own ‘--cnk_dmn’ options that will apply to all variables. cnk_plc key values: ‘r1d’, ‘cnk_r1d’, ‘plc_r1d’
Mnemonic: Record 1-D variables

Chunk Variables Containing Explicitly Chunked Dimensions

Definition: Chunk all variables possible that contain at least one dimension whose chunksize was explicitly set with the ‘--cnk_dmn’ option. Alternate invocation: none
cnk_plc key values: ‘xpl’, ‘cnk_xpl’, ‘plc_xpl’
Mnemonic: EXPLicitly specified dimensions

Chunk Variables that are already Chunked

Definition: Chunk only variables that are already chunked in the input file. When used in conjunction with ‘cnk_map=xst’ this option preserves and copies the chunking parameters from the input to the output file. Alternate invocation: none
cnk_plc key values: ‘xst’, ‘cnk_xst’, ‘plc_xst’
Mnemonic: EXiSTing chunked variables

Chunk Variables with NCO recommendations

Definition: Chunk all variables according to NCO best practices. This is a virtual option that ensures the chunking policy is (in the subjective opinion of the authors) the best policy for typical usage. As of NCO version 4.4.8 (February, 2015), this virtual policy implements ‘map_rew’ for 3-D variables and ‘map_lfp’ for all other variables.
Alternate invocation: none
cnk_plc key values: ‘nco’, ‘cnk_nco’, ‘plc_nco’
Mnemonic: NetCDFOperator

Unchunking

Definition: Unchunk all variables possible. The HDF5 storge layer requires that record variables (i.e., variables that contain at least one record dimension) must be chunked. Also variables that are compressed or use checksums must be chunked. Such variables cannot be unchunked.
Alternate invocation: ncunchunk
cnk_plc key values: ‘uck’, ‘cnk_uck’, ‘plc_uck’, ‘unchunk’
Mnemonic: UnChunK

Equivalent key values are fully interchangeable. Multiple equivalent options are provided to satisfy disparate needs and tastes of NCO users working with scripts and from the command line.

The chunking algorithms must know the chunksizes of each dimension of each variable to be chunked. The correspondence between the input variable shape and the chunksizes is called the chunking map. The user specifies the desired chunking map with the ‘-M’ switch (or its long option equivalents, ‘--cnk_map’ and ‘--chunk_map’) and its cnk_map argument. Nine chunking maps are currently implemented:

Chunksize Equals Dimension Size

Definition: Chunksize defaults to dimension size. Explicitly specify chunksizes for particular dimensions with ‘--cnk_dmn’ option.
cnk_map key values: ‘dmn’, ‘cnk_dmn’, ‘map_dmn’
Mnemonic: DiMeNsion

Chunksize Equals Dimension Size except Record Dimension

Definition: Chunksize equals dimension size except record dimension has size one. Explicitly specify chunksizes for particular dimensions with ‘--cnk_dmn’ option.
cnk_map key values: ‘rd1’, ‘cnk_rd1’, ‘map_rd1’
Mnemonic: Record Dimension size 1

Chunksize Equals Scalar Size Specified

Definition: Chunksize for all dimensions is set with the ‘--cnk_scl=sz_lmn’ option. For this map sz_lmn itself becomes the chunksize of each dimension. This is in contrast to the cnk_prd map, where the rth root of sz_lmn) becomes the chunksize of each dimension.
cnk_map key values: ‘scl’, ‘cnk_scl’, ‘map_scl’
Mnemonic: SCaLar
cnk_map key values: ‘xpl’, ‘cnk_xpl’, ‘map_xpl’
Mnemonic: EXPLicitly specified dimensions

Chunksize Product Matches Scalar Size Specified

Definition: The product of the chunksizes for each variable matches (approximately equals) the size specified with the ‘--cnk_scl=sz_lmn’ option. A dimension of size one is said to be degenerate. For a variable of rank R (i.e., with R non-degenerate dimensions), the chunksize in each non-degenerate dimension is (approximately) the Rth root of sz_lmn. This is in contrast to the cnk_scl map, where sz_lmn itself becomes the chunksize of each dimension.
cnk_map key values: ‘prd’, ‘cnk_prd’, ‘map_prd’
Mnemonic: PRoDuct

Chunksize Lefter Product Matches Scalar Size Specified

Definition: The product of the chunksizes for each variable (approximately) equals the size specified with the ‘--cnk_byt=sz_byt’ (not ‘--cnk_dfl’) option. This is accomplished by using dimension sizes as chunksizes for the rightmost (most rapidly varying) dimensions, and then “flexing” the chunksize of the leftmost (least rapidly varying) dimensions such that the product of all chunksizes matches the specified size. All L-dimensions to the left of and including the first record dimension define the left-hand side. To be precise, if the total size (in bytes) of the variable is var_sz, and if the specified (with ‘--cnk_byt’) product of the R “righter” dimensions (those that vary more rapidly than the first record dimension) is sz_byt, then chunksize (in bytes) of each of the L lefter dimensions is (approximately) the Lth root of var_sz/sz_byt. This map was first proposed by Chris Barker.
cnk_map key values: ‘lfp’, ‘cnk_lfp’, ‘map_lfp’
Mnemonic: LeFter Product

Chunksize Equals Existing Chunksize

Definition: Chunksizes are copied from the input to the output file for every variable that is chunked in the input file. Variables not chunked in the input file will be chunked with default mappings.
cnk_map key values: ‘xst’, ‘cnk_xst’, ‘map_xst’
Mnemonic: EXiST

Chunksize Balances 1D and (N-1)-D Access to N-D Variable [default for netCDF4 input]

Definition: Chunksizes are chosen so that 1-D and (N-1)-D hyperslabs of 3-D variables (e.g., point-timeseries orn latitude/longitude surfaces of 3-D fields) both require approximately the number of chunks. Hence their access time should be balanced. Russ Rew explains the motivation and derivation for this strategy here.
cnk_map key values: ‘rew’, ‘cnk_rew’, ‘map_rew’
Mnemonic: Russ REW

Chunksizes use netCDF4 defaults

Definition: Chunksizes are determined by the underlying netCDF library. All variables selected by the current chunking policy have their chunksizes determined by netCDF library defaults. The default algorithm netCDF uses to determine chunksizes has changed through the years, and thus depends on the netCDF library version. This map can be used to reset (portions of) previously chunked files to default chunking values.
cnk_map key values: ‘nc4’, ‘cnk_nc4’, ‘map_nc4’
Mnemonic: NetCDF4

Chunksizes use NCO recommendations [default for netCDF3 input]

Definition: Chunksizes are determined by the currently recommended NCO map. This is a virtual option that ensures the chunking map is (in the subjective opinion of the authors) the best map for typical usage. As of NCO version 4.4.9 (May, 2015), this virtual map calls ‘map_lfp’.
cnk_map key values: ‘nco’, ‘cnk_nco’, ‘map_nco’
Mnemonic: NetCDFOperator

It is possible to combine the above chunking map algorithms with user-specified per-dimension (though not per-variable) chunksizes that override specific chunksizes determined by the maps above. The user specifies the per-dimension chunksizes with the (equivalent) long options ‘--cnk_dmn’ or ‘--chunk_dimension’). The option takes two comma-separated arguments, dmn_nm,sz_lmn, which are the dimension name and its chunksize (in elements, not bytes), respectively. The ‘--cnk_dmn’ option may be used as many times as necessary.

The default behavior of chunking depends on several factors. As mentioned above, when no chunking options are explicitly specified by the user, then NCO preserves the chunking characteristics of the input file in the output file. This is equivalent to specifying both cnk_plc and cnk_map as “existing”, i.e., ‘--cnk_plc=xst --cnk_map=xst’. If output netCDF4 files are chunked with the default behavior of the netCDF4 library.

When any chunking parameter except ‘cnk_plc’ or ‘cnk_map’ is specified (such as ‘cnk_dmn’ or ‘cnk_scl’), then the “existing” policy and map are retained and the output chunksizes are modified where necessary in accord with the user-specified parameter. When ‘cnk_map’ is specified and ‘cnk_plc’ is not, then NCO picks (what it thinks is) the optimal chunking policy. This has always been policy ‘map_g2d’. When ‘cnk_plc’ is specified and ‘cnk_map’ is not, then NCO picks (what it thinks is) the optimal chunking map. This has always been map ‘map_rd1’.

To start afresh and return to netCDF4 chunking defaults, select ‘cnk_map=nc4’.

# Simple chunking and unchunking
ncks -O -4 --cnk_plc=all     in.nc out.nc # Chunk in.nc
ncks -O -4 --cnk_plc=unchunk in.nc out.nc # Unchunk in.nc

# Chunk data then unchunk it, printing informative metadata
ncks -O -4 -D 4 --cnk_plc=all ~/nco/data/in.nc ~/foo.nc
ncks -O -4 -D 4 --cnk_plc=uck ~/foo.nc ~/foo.nc

# Set total chunksize to 8192 B
ncks -O -4 -D 4 --cnk_plc=all --cnk_byt=8192 ~/nco/data/in.nc ~/foo.nc

# More complex chunking procedures, with informative metadata
ncks -O -4 -D 4 --cnk_scl=8 ~/nco/data/in.nc ~/foo.nc
ncks -O -4 -D 4 --cnk_scl=8 dstmch90_clm.nc ~/foo.nc
ncks -O -4 -D 4 --cnk_dmn lat,64 --cnk_dmn lon,128 dstmch90_clm.nc \ 
 ~/foo.nc 
ncks -O -4 -D 4 --cnk_plc=uck ~/foo.nc ~/foo.nc
ncks -O -4 -D 4 --cnk_plc=g2d --cnk_map=rd1 --cnk_dmn lat,32 \
 --cnk_dmn lon,128 dstmch90_clm_0112.nc ~/foo.nc

# Chunking works with all operators...
ncap2 -O -4 -D 4 --cnk_scl=8 -S ~/nco/data/ncap2_tst.nco \ 
 ~/nco/data/in.nc ~/foo.nc
ncbo -O -4 -D 4 --cnk_scl=8 -p ~/nco/data in.nc in.nc ~/foo.nc
ncecat -O -4 -D 4 -n 12,2,1 --cnk_dmn lat,32 \ 
 -p /data/zender/dstmch90 dstmch90_clm01.nc ~/foo.nc
ncflint -O -4 -D 4 --cnk_scl=8 ~/nco/data/in.nc ~/foo.nc
ncpdq -O -4 -D 4 -P all_new --cnk_scl=8 -L 5 ~/nco/data/in.nc ~/foo.nc
ncrcat -O -4 -D 4 -n 12,2,1 --cnk_dmn lat,32 \ 
 -p /data/zender/dstmch90 dstmch90_clm01.nc ~/foo.nc
ncwa -O -4 -D 4 -a time --cnk_plc=g2d --cnk_map=rd1 --cnk_dmn lat,32 \ 
 --cnk_dmn lon,128 dstmch90_clm_0112.nc ~/foo.nc

Chunking policy ‘r1d’ changes the chunksize of 1-D record variables (and no other variables) to that specified (with ‘--cnk_dmn’) chunksize. Any specified record dimension chunksizes will be applied to 1-D record variables only. Other dimensions may be chunked with their own ‘--cnk_dmn’ options that will apply to all variables. For example,

ncks --cnk_plc=r1d --cnk_dmn=time,1000. in.nc out.nc

This sets time chunks to 1000 only in 1-D record variables. Without the ‘r1d’ policy, time chunks would change in all variables.

It is appropriate to conclude by informing users about an aspect of chunking that may not be expected. Three types of variables are always chunked: Record variables, Deflated (compressed) variables, and Checksummed variables. Hence all variables that contain a record dimension are also chunked (since data must be chunked in all dimensions, not just one). Unless otherwise specified by the user, the other (fixed, non-record) dimensions of record variables are assigned default chunk sizes. The HDF5 layer does all this automatically to optimize the on-disk variable/file storage geometry of record variables. Do not be surprised to learn that files created without any explicit instructions to activate chunking nevertheless contain chunked variables.

3.29 Compression

NCO implements or accesses four different compression algorithms, the standard lossless DEFLATE algorithm and three lossy compression algorithms. All four algorithms reduce the on-disk size of a dataset while sacrificing no (lossless) or a tolerable amount (lossy) of precision. First, NCO can access the lossless DEFLATE algorithm, a combination of Lempel-Ziv encoding and Huffman coding, algorithm on any netCDF4 dataset (see Deflation). Because it is lossless, this algorithm re-inflates deflated data to their full original precision. This algorithm is accessed via the HDF5 library layer (which itself calls the zlib library also used by gzip), and is unavailable with netCDF3.

3.29.1 Linear Packing

The three lossy compression algorithms are Linear Packing (see Packed data), and two precision-preserving algorithms. Linear packing quantizes data of a higher precision type into a lower precision type (often NC_SHORT) that thus stores a fewer (though constant) number of bytes per value. Linearly packed data unpacks into a (much) smaller dynamic range than the floating-point data can represent. The type-conversion and reduced dynamic range of the data allows packing to eliminate bits typically used to store an exponent, thus improving its packing efficiency. Packed data also can also be deflated for additional space savings.

A limitation of linear packing is that unpacking data stored as integers into the linear range defined by scale_factor and add_offset rapidly loses precision outside of a narrow range of floating-point values. Variables packed as NC_SHORT, for example, can represent only about 64000 discrete values in the range -32768*scale_factor+add_offset to 32767*scale_factor+add_offset. The precision of packed data equals the value of scale_factor, and scale_factor is usually chosen to span the range of valid data, not to represent the intrinsic precision of the variable. In other words, the precision of packed data cannot be specified in advance because it depends on the range of values to quantize.

3.29.2 Precision-Preserving Compression

NCO implemented the final two lossy compression algorithms in version 4.4.8 (February, 2015). These are both Precision-Preserving Compression (PPC) algorithms and since standard terminology for precision is remarkably imprecise, so is our nomenclature. The operational definition of “significant digit” in our precision preserving algorithms is that the exact value, before rounding or quantization, is within one-half the value of the decimal place occupied by the Least Significant Digit (LSD) of the rounded value. For example, the value pi = 3.14 correctly represents the exact mathematical constant pi to three significant digits because the LSD of the rounded value (i.e., 4) is in the one-hundredths digit place, and the difference between the exact value and the rounded value is less than one-half of one one-hundredth, i.e., (3.14159265358979323844 - 3.14 = 0.00159 < 0.005).

One PPC algorithm preserves the specified total Number of Signifcant Digits (NSD) of the value. For example there is only one significant digit in the weight of most “eight-hundred pound gorillas” that you will encounter, i.e., so nsd=1. This is the most straightforward measure of precision, and thus NSD is the default PPC algorithm.

The other PPC algorithm preserves the number of Decimal Significant Digits (DSD), i.e., the number of significant digits following (positive, by convention) or preceding (negative) the decimal point. For example, ‘0.008’ and ‘800’ have, respectively, three and negative two digits digits following the decimal point, corresponding to dsd=3 and dsd=-2.

The only justifiable NSD for a given value depends on intrinsic accuracy and error characteristics of the model or measurements, and not on the units with which the value is stored. The appropriate DSD for a given value depends on these intrinsic characteristics and, in addition, the units of storage. This is the fundamental difference between the NSD and DSD approaches. The eight-hundred pound gorilla always has nsd=1 regardless of whether the value is stored in pounds or in some other unit. DSD corresponding to this weight is dsd=-2 if the value is stored in pounds, dsd=4 if stored in megapounds.

Users may wish to express the precision to be preserved as either NSD or DSD. Invoke PPC with the long option ‘--ppc var=prc’, or give the same arguments to the synonyms ‘--precision_preserving_compression’, or to ‘--quantize’. Here var is the variable to quantize, and prc is its precision. The option ‘--ppc’ (and its long option equivalents such as ‘--quantize’) indicates the argument syntax will be key=val. As such, ‘--ppc’ and its synonyms are indicator options that accept arguments supplied one-by-one like ‘--ppc key1=val1 --ppc key2=val2’, or aggregated together in multi-argument format like ‘--ppc key1=val1#key2=val2’ (see Multi-arguments). The default algorithm assumes prc specifies NSD precision, e.g., ‘T=2’ means nsd=2. Prepend prc with a decimal point to specify DSD precision, e.g., ‘T=.2’ means dsd=2. NSD precision must be specified as a positive integer. DSD precision may be a positive or negative integer; and is specified as the negative base 10 logarithm of the desired precision, in accord with common usage. For example, specifying ‘T=.3’ or ‘T=.-2’ tells the DSD algorithm to store only enough bits to preserve the value of T rounded to the nearest thousandth or hundred, respectively.

Setting var to default has the special meaning of applying the associated NSD or DSD algorithm to all floating point variables except coordinate variables. Variables not affected by default include integer and non-numeric atomic types, coordinates, and variables mentioned in the bounds, climatology, or coordinates attribute of any variable. NCO applies PPC to coordinate variables only if those variables are explicitly specified (i.e., not with the ‘default=prc’ mechanism. NCO applies PPC to integer-type variables only if those variables are explicitly specified (i.e., not with the ‘default=prc’, and only if the DSD algorithm is invoked with a negative prc. To prevent PPC from applying to certain non-coordinate variables (e.g., gridcell_area or gaussian_weight), explicitly specify a precision exceeding 7 (for NC_FLOAT) or 15 (for NC_DOUBLE) for those variables. Since these are the maximum representable precisions in decimal digits, NCO turns-off PPC (i.e., does nothing) when more precision is requested.

The time-penalty for compressing and uncompressing data varies according to the algorithm. The Number of Significant Digit (NSD) algorithm quantizes by bitmasking, and employs no floating-point math. The Decimal Significant Digit (DSD) algorithm quantizes by rounding, which does require floating-point math. Hence NSD is likely faster than DSD, though the difference has not been measured. NSD creates a bitmask to alter the significand of IEEE 754 floating-point data. The bitmask is one for all bits to be retained and zero or one for all bits to be ignored. The algorithm assumes that the number of binary digits (i.e., bits) necessary to represent a single base-10 digit is ln(10)/ln(2) = 3.32. The exact numbers of bits Nbit retained for single and double precision values are ceil(3.32*nsd)+1 and ceil(3.32*nsd)+2, respectively. Once these reach 23 and 53, respectively, bitmasking is completely ineffective. This occurs at nsd=6.3 and 15.4, respectively.

The DSD algorithm, by contrast, uses rounding to remove undesired precision. The rounding ⁴⁰ zeroes the greatest number of significand bits consistent with the desired precision.

To demonstrate the change in IEEE representation caused by PPC rounding algorithms, consider again the case of pi, represented as an NC_FLOAT. The IEEE 754 single precision representations of the exact value (3.141592...), the value with only three significant digits treated as exact (3.140000...), and the value as stored (3.140625) after PPC-rounding with either the NSD (prc=3) or DSD (prc=2) algorithm are, respectively,

S Exponent  Fraction (Significand)   Decimal    Notes
0 100000001 0010010000111111011011 # 3.14159265 Exact
0 100000001 0010001111010111000011 # 3.14000000
0 100000001 0010010000000000000000 # 3.14062500 NSD = 3
0 100000001 0010010000000000000000 # 3.14062500 DSD = 2

The string of trailing zero-bits in the rounded values facilitates byte-stream compression. Note that the NSD and DSD algorithms do not always produce results that are bit-for-bit identical, although they do in this particular case.

Reducing the preserved precision of NSD-rounding produces increasingly long strings of identical-bits amenable to compression:

S Exponent  Fraction (Significand)   Decimal    Notes
0 100000001 0010010000111111011011 # 3.14159265 Exact
0 100000001 0010010000111111011011 # 3.14159265 NSD = 8
0 100000001 0010010000111111011010 # 3.14159262 NSD = 7
0 100000001 0010010000111111011000 # 3.14159203 NSD = 6
0 100000001 0010010000111111000000 # 3.14158630 NSD = 5
0 100000001 0010010000111100000000 # 3.14154053 NSD = 4
0 100000001 0010010000000000000000 # 3.14062500 NSD = 3
0 100000001 0010010000000000000000 # 3.14062500 NSD = 2
0 100000001 0010000000000000000000 # 3.12500000 NSD = 1

The consumption of about 3 bits per digit of base-10 precision is evident, as is the coincidence of a quantized value that greatly exceeds the mandated precision for NSD = 2. Although the NSD algorithm generally masks some bits for all nsd <= 7 (for NC_FLOAT), compression algorithms like DEFLATE may need byte-size-or-greater (i.e., at least eight-bit) bit patterns before their algorithms can take advantage of of encoding such patterns for compression. Do not expect significantly enhanced compression from nsd > 5 (for NC_FLOAT) or nsd > 14 (for NC_DOUBLE). Clearly values stored as NC_DOUBLE (i.e., eight-bytes) are susceptible to much greater compression than NC_FLOAT for a given precision because their significands explicitly contain 53 bits rather than 23 bits.

Maintaining non-biased statistical properties during lossy compression requires special attention. The DSD algorithm uses rint(), which rounds toward the nearest even integer. Thus DSD has no systematic bias. However, the NSD algorithm uses a bitmask technique susceptible to statistical bias. Zeroing all non-significant bits is guaranteed to produce numbers quantized to the specified tolerance, i.e., half of the decimal value of the position occupied by the LSD. However, always zeroing the non-significant bits results in quantized numbers that never exceed the exact number. This would produce a negative bias in statistical quantities (e.g., the average) subsequently derived from the quantized numbers. To avoid this bias, our NSD implementation rounds non-significant bits down (to zero) or up (to one) in an alternating fashion when processing array data. In general, the first element is rounded down, the second up, and so on. This results in a mean bias quite close to zero. The only exception is that the floating-point value of zero is never quantized upwards. For simplicity, NSD always rounds scalars downwards.

Although NSD or DSD are different algorithms under the hood, they both replace the (unwanted) least siginificant bits of the IEEE significand with a string of consecutive zeroes. Byte-stream compression techniques, such as the gzip DEFLATE algorithm compression available in HDF5, always compress zero-strings more efficiently than random digits. The net result is netCDF files that utilize compression can be significantly reduced in size. This feature only works when the data are compressed, either internally (by netCDF) or externally (by another user-supplied mechanism). It is most straightfoward to compress data internally using the built-in compression and decompression supported by netCDF4. For convenience, NCO automatically activates file-wide Lempel-Ziv deflation (see Deflation) level one (i.e., ‘-L 1’) when PPC is invoked on any variable in a netCDF4 output file. This makes PPC easier to use effectively, since the user need not explicitly specify deflation. Any explicitly specified deflation (including no deflation, ‘-L 0’) will override the PPC deflation default. If the output file is a netCDF3 format, NCO will emit a message suggesting internal netCDF4 or external netCDF3 compression. netCDF3 files compressed by an external utility such as gzip accrue approximately the same benefits (shrinkage) as netCDF4, although with netCDF3 the user or provider must uncompress (e.g., gunzip) the file before accessing the data. There is no benefit to rounding numbers and storing them in netCDF3 files unless such custom compression/decompression is employed. Without that, one may as well maintain the undesired precision.

The user accesses PPC through a single switch, ‘--ppc’, repeated as many times as necessary. To apply the NSD algorithm to variable u use, e.g.,

ncks -7 --ppc u=2 in.nc out.nc

The output file will preserve only two significant digits of u. The options ‘-4’ or ‘-7’ ensure a netCDF4-format output (regardless of the input file format) to support internal compression. It is recommended though not required to write netCDF4 files after PPC. For clarity the ‘-4/-7’ switches are omitted in subsequent examples. NCO attaches attributes that indicate the algorithm used and degree of precision retained for each variable affected by PPC. The NSD and DSD algorithms store the attributes number_of_significant_digits and least_significant_digit ⁴¹, respectively.

It is safe to attempt PPC on input that has already been rounded. Variables can be made rounder, not sharper, i.e., variables cannot be “un-rounded”. Thus PPC attempted on an input variable with an existing PPC attribute proceeds only if the new rounding level exceeds the old, otherwise no new rounding occurs (i.e., a “no-op”), and the original PPC attribute is retained rather than replaced with the newer value of prc.

To request, say, five significant digits (nsd=5) for all fields, except, say, wind speeds which are only known to integer values (dsd=0) in the supplied units, requires ‘--ppc’ twice:

ncks -4 --ppc default=5 --ppc u,v=.0 in.nc out.nc

To preserve five digits in all variables except coordinate variables and u and v, use the ‘default’ option and separately specify the exceptions:

ncks --ppc default=5 --ppc u,v=20 in.nc out.nc

The ‘--ppc’ option may be specified any number of times to support varying precision types and levels, and each option may aggregate all the variables with the same precision

ncks --ppc p,w,z=5 --ppc q,RH=4 --ppc T,u,v=3 in.nc out.nc
ncks --ppc p,w,z=5#q,RH=4#T,u,v=3 in.nc out.nc # Multi-argument format

Any var argument may be a regular expression. This simplifies generating lists of related variables:

ncks --ppc Q.?=5 --ppc FS.?,FL.?=4 --ppc RH=.3 in.nc out.nc
ncks --ppc Q.?=5#FS.?,FL.?=4#RH=.3 in.nc out.nc # Multi-argument format

Although PPC-rounding instantly reduces data precision, on-disk storage reduction only occurs once the data are compressed.

How can one be sure the lossy data are sufficiently precise? PPC preserves all significant digits of every value. The DSD algorithm uses floating-point math to round each value optimally so that it has the maximum number of zeroed bits that preserve the specified precision. The NSD algorithm uses a theoretical approach (3.2 bits per base-10 digit), tuned and tested to ensure the worst case quantization error is less than half the value of the minimum increment in the least significant digit.

All three metrics are expressed in terms of the fraction of the ten’s place occupied by the LSD. If the LSD is the hundreds digit or the thousandths digit, then the metrics are fractions of 100, or of 1/100, respectively. PPC algorithms should produce maximum absolute errors no greater than 0.5 in these units. If the LSD is the hundreds digit, then quantized versions of true values will be within fifty of the true value. It is much easier to satisfy this tolerance for a true value of 100 (only 50% accuracy required) than for 999 (5% accuracy required). Thus the minimum accuracy guaranteed for nsd=1 ranges from 5–50%. For this reason, the best and worst cast performance usually occurs for true values whose LSD value is close to one and nine, respectively. Of course most users prefer prc > 1 because accuracies increase exponentially with prc. Continuing the previous example to prc=2, quantized versions of true values from 1000–9999 will also be within 50 of the true value, i.e., have accuracies from 0.5–5%. In other words, only two significant digits are necessary to guarantee better than 5% accuracy in quantization. We recommend that dataset producers and users consider quantizing datasets with nsd=3. This guarantees accuracy of 0.05–0.5% for individual values. Statistics computed from ensembles of quantized values will, assuming the mean error Emean is small, have much better accuracy than 0.5%. This accuracy is the most that can be justified for many applications.

To demonstrate these principles we conduct error analyses on an artificial, reproducible dataset, and on an actual dataset of observational analysis values. ⁴² The table summarizes quantization accuracy based on the three metrics.

NSD

Number of Significant Digits.

Emabs

Maximum absolute error.

Emebs

Mean absolute error.

Emean

Mean error.

Artificial Data: N=1000000 values in [1.0,2.0) in steps of 1.0e-6
Single-Precision        Double-Precision   Single-Precision
NSD Emabs Emebs Emean   Emabs Emebs Emean  DSD Emabs Emebs Emean
 1  0.31  0.11  4.1e-4  0.31  0.11  4.0e-4  1  0.30  0.11 -8.1e-4  
 2  0.39  0.14  6.8e-5  0.39  0.14  5.5e-5  2  0.39  0.14 -1.3e-4
 3  0.49  0.17  1.0e-6  0.49  0.17 -5.5e-7  3  0.49  0.17 -2.0e-5
 4  0.30  0.11  3.2e-7  0.30  0.11 -6.1e-6  4  0.30  0.11  5.1e-8
 5  0.37  0.13  3.1e-7  0.38  0.13 -5.6e-6  5  0.38  0.13  2.6e-6
 6  0.36  0.12 -4.4e-7  0.48  0.17 -4.1e-7  6  0.48  0.17  7.2e-6
 7  0.00  0.00  0.0     0.30  0.10  1.5e-7  7  0.00  0.00  0.0     

Observational Analysis: N=13934592 values MERRA Temperature 20130601
Single-Precision        
NSD Emabs Emebs Emean   
 1  0.31  0.11  2.4e-3
 2  0.39  0.14  3.8e-4
 3  0.49  0.17 -9.6e-5 
 4  0.30  0.11  2.3e-3
 5  0.37  0.13  2.2e-3
 6  0.36  0.13  1.7e-2
 7  0.00  0.00  0.0     

All results show that PPC quantization performs as expected. Absolute maximum errors Emabs < 0.5 for all prc. For 1 <= prc <= 6, quantization results in comparable maximum absolute and mean absolute errors Emabs and Emebs, respectively. Mean errors Emean are orders of magnitude smaller because quantization produces over- and under-estimated values in balance. When prc=7, quantization of single-precision values is ineffective, because all available bits are used to represent the maximum precision of seven digits. The maximum and mean absolute errors Emabs and Emebs are nearly identical across algorithms, precisions, and dataset types. This is consistent with both the artificial data and empirical data being random, and thus exercising equally strengths and weaknesses of the algorithms over the course of millions of input values. We generated artificial arrays with many different starting values and interval spacing and all gave qualitatively similar results. The results presented are the worst obtained.

The artificial data has much smaller mean error Emean than the observational analysis. The reason why is unclear. It may be because the temperature field is concentrated in particular ranges of values (and associated quantization errors) prevalent on Earth, e.g., 200 < T < 320. It is worth noting that the mean error Emean < 0.01 for 1 <= prc < 6, and that Emean is typically at least two or more orders of magnitude less than Emabs. Thus quantized values with precisions as low as prc=1 still yield highly significant statistics by contemporary scientific standards.

Testing shows that PPC quantization enhances compression of typical climate datasets. The degree of enhancement depends, of course, on the required precision. Model results are often computed as NC_DOUBLE then archived as NC_FLOAT to save space. This table summarizes the performance of lossless and lossy compression on two typical, or at least random, netCDF data files. The files were taken from representative model-simulated and satellite-retrieved datasets. Only floating-point data were compressed. No attempt was made to compress integer-type variables as they occupy an insignificant fraction of every dataset. The columns are

Type

File-type: N3 for netCDF CLASSIC, N4 for NETCDF4, N7 for NETCDF4_CLASSIC (which comprises netCDF3 data types and structures with netCDF4 storage features like compression), H4 for HDF4, and H5 for HDF5. N4/7 means results apply to both N4 and N7 filetypes.

LLC

Type of lossless compression employed, if any. Bare numbers refer to the strength of the DEFLATE algorithm employed internally by netCDF4/HDF5, while numbers prefixed with B refer to the block size employed by the Burrows-Wheeler algorithm in bzip2.

PPC

Number of significant digits retained by the precision-preserving compression NSD algorithm.

Pck

Y if the default ncpdq packing algorithm (convert floating-point types to NC_SHORT) was employed.

Size

Resulting filesize in MB.

%

Compression ratio, i.e., resulting filesize relative to original size, in percent. In some cases the original files is already losslessly compressed. The compression ratios reported are relative to the size of the original file as distributed, not as optimally losslessly compressed.

A dash (-) indicates the associated compression feature was not employed.

# dstmch90_clm.nc
Type LLC PPC Pck  Size   %    Flags and Notes
  N3   -   -  -   34.7 100.0  Original is not compressed
  N3  B1   -  -   28.9  83.2  bzip2 -1
  N3  B9   -  -   29.3  84.4  bzip2 -9
  N7   -   -  -   35.0 101.0     
  N7   1   -  -   28.2  81.3  -L 1
  N7   9   -  -   28.0  80.8  -L 9
  N7   -   -  Y   17.6  50.9  ncpdq -L 0
  N7   1   -  Y    7.9  22.8  ncpdq -L 1
  N7   1   7  -   28.2  81.3  --ppc default=7
  N7   1   6  -   27.9  80.6  --ppc default=6
  N7   1   5  -   25.9  74.6  --ppc default=5
  N7   1   4  -   22.3  64.3  --ppc default=4
  N7   1   3  -   18.9  54.6  --ppc default=3
  N7   1   2  -   14.5  43.2  --ppc default=2
  N7   1   1  -   10.0  29.0  --ppc default=1

# b1850c5cn_doe_polar_merged_0_cesm1_2_0_HD+MAM4+tun2b.hp.e003.cam.h0.0001-01.nc
Type LLC PPC Pck  Size   %    Flags and Notes
  N3   -   -  -  119.8 100.0  Original is not compressed
  N3  B1   -  -   84.2  70.3  bzip2 -1
  N3  B9   -  -   84.8  70.9  bzip2 -9
  N7   -   -  -  120.5 100.7     
  N7   1   -  -   82.6  69.0  -L 1
  N7   9   -  -   82.1  68.6  -L 9
  N7   -   -  Y   60.7  50.7  ncpdq -L 0
  N7   1   -  Y   26.0  21.8  ncpdq -L 1
  N7   1   7  -   82.6  69.0  --ppc default=7
  N7   1   6  -   81.9  68.4  --ppc default=6
  N7   1   5  -   77.2  64.5  --ppc default=5
  N7   1   4  -   69.0  57.6  --ppc default=4
  N7   1   3  -   59.3  49.5  --ppc default=3
  N7   1   2  -   49.5  41.3  --ppc default=2
  N7   1   1  -   38.2  31.9  --ppc default=1

# MERRA300.prod.assim.inst3_3d_asm_Cp.20130601.hdf
Type LLC PPC Pck  Size   %    Flags and Notes
  H4   5   -  -  244.3 100.0  Original is compressed
  H4  B1   -  -  244.7 100.1  bzip2 -1
  N4   5   -  -  214.5  87.8
  N7   5   -  -  210.6  86.2  
  N4  B1   -  -  215.4  88.2  bzip2 -1
  N4  B9   -  -  214.8  87.9  bzip2 -9
  N3   -   -  -  617.1 252.6
N4/7   -   -  -  694.0 284.0  -L 0
N4/7   1   -  -  223.2  91.3  -L 1
N4/7   9   -  -  207.3  84.9  -L 9
N4/7   -   -  Y  347.1 142.1  ncpdq -L 0
N4/7   1   -  Y  133.6  54.7  ncpdq -L 1
N4/7   1   7  -  223.1  91.3  --ppc default=7
N4/7   1   6  -  225.1  92.1  --ppc default=6
N4/7   1   5  -  221.4  90.6  --ppc default=5
N4/7   1   4  -  201.4  82.4  --ppc default=4
N4/7   1   3  -  185.3  75.9  --ppc default=3
N4/7   1   2  -  150.0  61.4  --ppc default=2
N4/7   1   1  -  100.8  41.3  --ppc default=1

# OMI-Aura_L2-OMIAuraSO2_2012m1222-o44888_v01-00-2014m0107t114720.h5
Type LLC PPC Pck  Size   %    Flags and Notes
  H5   5   -  -   29.5 100.0  Original is compressed
  H5  B1   -  -   29.3  99.6  bzip2 -1
  N4   5   -  -   29.5 100.0
  N4  B1   -  -   29.3  99.6  bzip2 -1
  N4  B9   -  -   29.3  99.4  bzip2 -9
  N4   -   -  -   50.7 172.3  -L 0
  N4   1   -  -   29.8 101.3  -L 1
  N4   9   -  -   29.4  99.8  -L 9
  N4   -   -  Y   27.7  94.0  ncpdq -L 0
  N4   1   -  Y   12.9  43.9  ncpdq -L 1
  N4   1   7  -   29.7 100.7  --ppc default=7
  N4   1   6  -   29.7 100.8  --ppc default=6
  N4   1   5  -   27.3  92.8  --ppc default=5
  N4   1   4  -   23.8  80.7  --ppc default=4
  N4   1   3  -   20.3  69.0  --ppc default=3
  N4   1   2  -   15.1  51.2  --ppc default=2
  N4   1   1  -    9.9  33.6  --ppc default=1

A selective, per-variable approach to PPC yields the best balance of precision and compression yet requires the dataset producer to understand the intrinsic precision of each variable. Such a specification for a GCM dataset might look like this (using names for the NCAR CAM model):

# Be conservative on non-explicit quantities, so default=5
# Some quantities deserve four significant digits
# Many quantities, such as aerosol optical depths and burdens, are 
# highly uncertain and only useful to three significant digits.
ncks -7 -O \
--ppc default=5 \
--ppc AN.?,AQ.?=4 \
--ppc AER.?,AOD.?,ARE.?,AW.?,BURDEN.?=3 \
ncar_cam.nc ~/foo.nc

3.30 Deflation

All NCO operators that define variables support the netCDF4 feature of storing variables compressed with the lossless DEFLATE compression algorithm. DEFLATE combines the Lempel-Ziv encoding with Huffman coding. The specific version used by netCDF4/HDF5 is that implemented in the zlib library used by gzip. Activate deflation with the -L dfl_lvl short option (or with the same argument to the ‘--dfl_lvl’ or ‘--deflate’ long options). Specify the deflation level dfl_lvl on a scale from no deflation (dfl_lvl = 0) to maximum deflation (dfl_lvl = 9). Under the hood, this selects the compression blocksize. Minimal deflation (dfl_lvl = 1) achieves considerable storage compression with little time penalty. Higher deflation levels require more time for compression. File sizes resulting from minimal (dfl_lvl = 1) and maximal (dfl_lvl = 9) deflation levels typically differ by less than 10% in size.

To compress an entire file using deflation, use

ncks -4 -L 0 in.nc out.nc # No deflation (fast, no time penalty)
ncks -4 -L 1 in.nc out.nc # Minimal deflation (little time penalty)
ncks -4 -L 9 in.nc out.nc # Maximal deflation (much slower)

Unscientific testing shows that deflation compresses typical climate datasets by 30-60%. Packing, a lossy compression technique available for all netCDF files (see Packed data), can easily compress files by 50%. Packed data may be deflated to squeeze datasets by about 80%:

ncks  -4 -L 1 in.nc out.nc # Minimal deflation (~30-60% compression)
ncks  -4 -L 9 in.nc out.nc # Maximal deflation (~31-63% compression)
ncpdq         in.nc out.nc # Standard packing  (~50% compression)
ncpdq -4 -L 9 in.nc out.nc # Deflated packing  (~80% compression)

ncks prints deflation parameters, if any, to screen (see ncks netCDF Kitchen Sink).

3.31 MD5 digests

As of NCO version 4.1.0 (April, 2012), NCO supports data integrity verification using the MD5 digest algorithm. This support is currently implemented in ncks and in the multifile concantenators ncecat and ncrcat. Activate it with the ‘--md5_dgs’ or ‘--md5_digest’ long options. As of NCO version 4.3.3 (July, 2013), NCO will write the MD5 digest of each variable as an NC_CHAR attribute named MD5. This support is currently implemented in ncks and in the multifile concantenators ncecat and ncrcat. Activate it with the ‘--md5_wrt_att’ or ‘--md5_write_attribute’ long options.

The behavior and verbosity of the MD5 digest is operator-dependent. When activating MD5 digests with ncks it is assumed that the user simply wishes to see the digest of every variable and this is done when the debugging level exceeds one. This incurs only the minor overhead of performing the hash algorithm for each variable read. MD5 digests may be activated in both the one- and two-filename argument forms of ncks, which are used for printing and for sub-setting, respectively. The MD5 digests are shown as a 32-character hexadecimal string in which each two characters represent one byte of the 16-byte digest:

> ncks -O -D 2 -C --md5 -v md5_a,md5_abc ~/nco/data/in.nc
...
ncks: INFO MD5(md5_a) = 0cc175b9c0f1b6a831c399e269772661
md5_a = 'a' 
ncks: INFO MD5(md5_abc) = 900150983cd24fb0d6963f7d28e17f72
lev[0]=100 md5_abc[0--2]='abc' 
> ncks -O -D 2 -C -d lev,0 --md5 -v md5_a,md5_abc ~/nco/data/in.nc
...
ncks: INFO MD5(md5_a) = 0cc175b9c0f1b6a831c399e269772661
md5_a = 'a' 
ncks: INFO MD5(md5_abc) = 0cc175b9c0f1b6a831c399e269772661
lev[0]=100 md5_abc[0--0]='a' 

In fact these examples demonstrate the validity of the hash algorithm since the MD5 hashes of the strings “a” and “abc” are widely known. The second example shows that the hyperslab of variable md5_abc (= “abc”) consisting of only its first letter (= “a”) has the same hash as the variable md5_a (“a”). This illustrates that MD5 digests act only on variable data, not on metadata.

When activating MD5 digests with ncecat or ncrcat it is assumed that the user wishes to verify that every variable written to disk has the same MD5 digest as when it is subsequently read from disk. This incurs the major additional overhead of reading in each variable after it is written and performing the hash algorithm again on that to compare to the original hash. Moreover, it is assumed that such operations are generally done “production mode” where the user is not interested in actually examining the digests herself. The digests proceed silently unless the debugging level exceeds three:

> ncecat -O -D 4 --md5 -p ~/nco/data in.nc in.nc ~/foo.nc | grep MD5
...
ncecat: INFO MD5(wnd_spd) = bec190dd944f2ce2794a7a4abf224b28
ncecat: INFO MD5 digests of RAM and disk contents for wnd_spd agree
> ncrcat -O -D 4 --md5 -p ~/nco/data in.nc in.nc ~/foo.nc | grep MD5
...
ncrcat: INFO MD5(wnd_spd) = 74699bb0a72b7f16456badb2c995f1a1
ncrcat: INFO MD5 digests of RAM and disk contents for wnd_spd agree

Regardless of the debugging level, an error is returned when the digests of the variable read from the source file and from the output file disagree.

These rules are evolving and as NCO pays more attention to data integrity. We welcome feedback and suggestions from users.

3.32 Buffer sizes

As of NCO version 4.2.0 (May, 2012), NCO allows the user to request specific buffer sizes to allocate for reading and writing files. This buffer size determines how many system calls the netCDF layer must invoke to read and write files. By default, netCDF uses the preferred I/O block size returned as the ‘st_blksize’ member of the ‘stat’ structure returned by the stat() system call ⁴³. Otherwise, netCDF uses twice the system pagesize. Larger sizes can increase access speed by reducing the number of system calls netCDF makes to read/write data from/to disk. Because netCDF cannot guarantee the buffer size request will be met, the actual buffer size granted by the system is printed as an INFO statement.

# Request 2 MB file buffer instead of default 8 kB buffer
> ncks -O -D 3 --bfr_sz=2097152 ~/nco/data/in.nc ~/foo.nc
...
ncks: INFO nc__open() will request file buffer size = 2097152 bytes
ncks: INFO nc__open() opened file with buffer size = 2097152 bytes
...

3.33 RAM disks

As of NCO version 4.2.1 (August, 2012), NCO supports the use of diskless files, aka RAM disks, for file access and creation. Two independent switches, ‘--open_ram’ and ‘--create_ram’, control this feature. Before describing the specifics of these switches, we describe why many NCO operations will not benefit from them. Essentially, reading/writing from/to RAM rather than disk only hastens the task when reads/writes to disk are avoided. Most NCO operations are simple enough that they require a single read-from/write-to disk for every block of input/output. Diskless access does not change this, but it does add an extra read-from/write-to RAM. However this extra RAM write/read does avoid contention for limited system resources like disk-head access. Operators which may benefit from RAM disks include ncwa, which may need to read weighting variables multiple times, the multi-file operators ncra, ncrcat, and ncecat, which may try to write output at least once per input file, and ncap2 scripts which may be arbitrarily long and convoluted.

The ‘--open_ram’ switch causes input files to copied to RAM when opened. All further metadata and data access occurs in RAM and thus avoids access time delays caused by disk-head movement. Usually input data is read at most once so it is unlikely that requesting input files be stored in RAM will save much time. The likeliest exceptions are files that are accessed numerous times, such as those analyzed extensively analyzed by ncap2.

Invoking ‘--open_ram’, ‘--ram_all’, or ‘--diskless_all’ uses much more system memory. To copy the input file to RAM increases the sustained memory use by exactly the on-disk filesize of the input file, i.e., MS += FT. For large input files this can be a huge memory burden that starves the rest of the NCO analysis of sufficient RAM. To be safe, use ‘--open_ram’, ‘--ram_all’, or ‘--diskless_all’ only on files that are much (say at least a factor of four) smaller than your available system RAM. See Memory Requirements for further details.