$Header$ -*-text-*-

netCDF Operators NCO version 5.3.9 sets sail

http://nco.sf.net (Homepage, Mailing lists, Help)
http://github.com/nco/nco (Source Code, Issues, Releases)

What's new?
Version 5.3.9 contains one major improvement and one bugfix.
The default buffer size for all operators on classic files is 32 MiB.
This will increase I/O speed in most cases. ncremap had a syntax
error in a rarely used branch of code that caused Bash to emit scary
messages even when the branch was not executed.

Users who want more speed without have to change any commands should
upgrade to 5.3.9. 

Enjoy,
Charlie

NEW FEATURES (full details always in ChangeLog):

A. The default buffer size for operating on classic files is 32 MiB.
Formerly NCO used the netCDF default which was 8 KiB unless the
filesystem explicitly set a blocksize in the stat->st_blksize element,
in which case netCDF would set the buffersize to twice the blocksize.
Experience shows that too many filesystems do not set the blocksize
element and thus the blocksize remained (often) 8 KiB, which 
is much too small for efficient I/O on most ESM datasets. Hence we
changed it to 32 MiB, the same buffer size that netCDF4 starts with.
To return to the netCDF default buffer size, use --bfr_sz=0.
To change the buffer size set it to your favorite number. This buffer
size affects classic files only, it is ignored and harmless when set
for netCDF4 files. Performance can significantly improve with sizes up
to 1 GiB, though it is dependent on the filesystem, RAM, and the
sizes of the variables being processed. Thanks to Mark Taylor (ANL)
and Noel Keen (LBNL/NERSC) for help identifying this issue.

ncks in.nc out.nc # NCO default 32 MiB buffer size
ncks --bfr_sz=0 in.nc out.nc # Use netCDF default size
ncclimo --bfr_sz=1073741824 in.nc out.nc # Request 1 GiB buffer size

B. ncks -k (or --prn_fmt, --prn_fl_fmt) now prints the file format.
This is similar (by intention) to the ncdump -k command. The first
word ncks prints in response is identical to the ncdump -k response.
In addition, ncks also prints all the synonyms for that file type
(since clear naming is not a quality of netCDF filetypes), and the
capacity of that filetype to hold variables and types:

% ncks -k in.nc
classic = NC_FORMAT_CLASSIC = CDF1. This is the earliest ...

http://nco.sf.net/nco.html#prn_fmt

BUG FIXES:

A. ncremap had a syntax error in a rarely used branch of code that
caused Bash to emit scary messages even when the branch was not
executed. The branch could only be encountered when running in MPI
mode on machines that were not in ncremap's internal database.
There is no workaround, the fix is to upgrade.

Full release statement at http://nco.sf.net/ANNOUNCE
    
KNOWN PROBLEMS DUE TO NCO:

This section of ANNOUNCE reports and reminds users of the
existence and severity of known, not yet fixed, problems. 
These problems occur with NCO 5.3.9 built/tested under
MacOS 26.3.1 with netCDF 4.10.1-development on HDF5 2.1.1
and with Linux FC42 with netCDF 4.9.2 on HDF5 1.14.4.

A. NOT YET FIXED (NCO problem)
   Correctly read arrays of NC_STRING with embedded delimiters in ncatted arguments

   Demonstration:
   ncatted -D 5 -O -a new_string_att,att_var,c,sng,"list","of","str,ings" ~/nco/data/in_4.nc ~/foo.nc
   ncks -m -C -v att_var ~/foo.nc

   20130724: Verified problem still exists
   TODO nco1102
   Cause: NCO parsing of ncatted arguments is not sophisticated
   enough to handle arrays of NC_STRINGS with embedded delimiters.

B. NOT YET FIXED (NCO problem?)
   ncra/ncrcat (not ncks) hyperslabbing can fail on variables with multiple record dimensions

   Demonstration:
   ncrcat -O -d time,0 ~/nco/data/mrd.nc ~/foo.nc

   20140826: Verified problem still exists
   20140619: Problem reported by rmla
   Cause: Unsure. Maybe ncra.c loop structure not amenable to MRD?
   Workaround: Convert to fixed dimensions then hyperslab

KNOWN PROBLEMS DUE TO BASE LIBRARIES/PROTOCOLS:

A. NOT YET FIXED (netCDF4 or HDF5 problem?)
   Specifying strided hyperslab on large netCDF4 datasets leads
   to slowdown or failure with recent netCDF versions.

   Demonstration with NCO <= 4.4.5:
   time ncks -O -d time,0,,12 ~/ET_2000-01_2001-12.nc ~/foo.nc
   Demonstration with NCL:
   time ncl < ~/nco/data/ncl.ncl   
   20140718: Problem reported by Parker Norton
   20140826: Verified problem still exists
   20140930: Finish NCO workaround for problem
   20190201: Possibly this problem was fixed in netCDF 4.6.2 by https://github.com/Unidata/netcdf-c/pull/1001
   Cause: Slow algorithm in nc_var_gets()?
   Workaround #1: Use NCO 4.4.6 or later (avoids nc_var_gets())
   Workaround #2: Convert file to netCDF3 first, then use stride
   Workaround #3: Compile NCO with netCDF >= 4.6.2

B. NOT YET FIXED (netCDF4 library bug)
   Simultaneously renaming multiple dimensions in netCDF4 file can corrupt output

   Demonstration:
   ncrename -O -d lev,z -d lat,y -d lon,x ~/nco/data/in_grp.nc ~/foo.nc # Completes but produces unreadable file foo.nc
   ncks -v one ~/foo.nc

   20150922: Confirmed problem reported by Isabelle Dast, reported to Unidata
   20150924: Unidata confirmed problem
   20160212: Verified problem still exists in netCDF library
   20160512: Ditto
   20161028: Verified problem still exists with netCDF 4.4.1
   20170323: Verified problem still exists with netCDF 4.4.2-development
   20170323: https://github.com/Unidata/netcdf-c/issues/381
   20171102: Verified problem still exists with netCDF 4.5.1-development
   20171107: https://github.com/Unidata/netcdf-c/issues/597
   20190202: Progress has recently been made in netCDF 4.6.3-development
   More details: http://nco.sf.net/nco.html#ncrename_crd

C. NOT YET FIXED (would require DAP protocol change?)
   Unable to retrieve contents of variables including period '.' in name
   Periods are legal characters in netCDF variable names.
   Metadata are returned successfully, data are not.
   DAP non-transparency: Works locally, fails through DAP server.

   Demonstration:
   ncks -O -C -D 3 -v var_nm.dot -p http://thredds-test.ucar.edu/thredds/dodsC/testdods in.nc # Fails to find variable

   20130724: Verified problem still exists. 
   Stopped testing because inclusion of var_nm.dot broke all test scripts.
   NB: Hard to fix since DAP interprets '.' as structure delimiter in HTTP query string.

   Bug tracking: https://www.unidata.ucar.edu/jira/browse/NCF-47

D. NOT YET FIXED (would require DAP protocol change)
   Correctly read scalar characters over DAP.
   DAP non-transparency: Works locally, fails through DAP server.
   Problem, IMHO, is with DAP definition/protocol

   Demonstration:
   ncks -O -D 1 -H -C -m --md5_dgs -v md5_a -p http://thredds-test.ucar.edu/thredds/dodsC/testdods in.nc

   20120801: Verified problem still exists
   Bug report not filed
   Cause: DAP translates scalar characters into 64-element (this
   dimension is user-configurable, but still...), NUL-terminated
   strings so MD5 agreement fails 

"Sticky" reminders:

A. Reminder that NCO works on most HDF4 and HDF5 datasets, e.g., 
   HDF4: AMSR MERRA MODIS ...
   HDF5: GLAS ICESat Mabel SBUV ...
   HDF-EOS5: AURA HIRDLS OMI ...

B. Pre-built executables for many OS's at:
   http://nco.sf.net#bnr

