Enabling netcdf deflation in CMOR output files

[durack1]

I am hoping to use the default netcdf4 deflation in output files written by CMOR. Is this possible using the python interface?

[durack1]

@dnadeau4 following our discussion today here are the discussion points:

• What should defaults be, speed of read vs file sizes?
• Should shuffling and chunking be enabled by default?
• How should a user specify deflation (in the CMIP6_Amon.json tables) or in the CMOR3 driver file (my preference is driver file, the tables should only contain information not runtime options for CMOR3)

We'll need to consult with @taylor13 to finalize the default behavior of CMOR3

[durack1]

@doutriaux1 has comments about this - regarding "standards". Long story short, a user should be able to over-ride the Table default at runtime, however including the attributes deflate, deflate_level and shuffle in the Tables ensures that across variables the same default compression is applied regardless of whether the CMOR user is aware of the deflation functionality or not.

[taylor13]

Balaji is strongly advising for us to check out performance issues. GFDL
has all but ruled out any sort of compression. We'll talk Wednesday.

On 6/6/16 12:40 PM, Paul J. Durack wrote:

@doutriaux1 https://github.com/doutriaux1 has comments about this -
regarding "standards". Long story short, a user should be able to
over-ride the Table default at runtime, however including the
attributes |deflate|, |deflate_level| and |shuffle| in the Tables
ensures that across variables the same default compression is applied
regardless of whether the CMOR user is aware of the deflation
functionality or not.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#57 (comment), or
mute the thread
https://github.com/notifications/unsubscribe/AEwxn9qRjJFGX3eaq_1qtoZQCNIpUh3Fks5qJHefgaJpZM4Iqtd1.

[durack1]

@dnadeau4 has this implemented in 3.0.6 with the following (python) notation:

cmor.set_deflate(varid,1,1,1) ; # shuffle=1,deflate=1,deflate_level=1 ; CMOR 3.0.6+
cmor.write( ...

[dnadeau4]

You can put the shuffle/deflation flag in the tables or call cmor.set_deflate. I have also add the API to C and Fortran. By default everything is netcdf3, I would like to default to netcdf4 with no compression.

[durack1]

@dnadeau4 I think defaulting to netcdf4 is a good idea - it is the requested format for CMIP6 as far as I know

[durack1]

Just FYI, here is what the file sizes look like using various deflate_level (level=X) flags with the AMIP boundary forcing datasets (1x1 lat/lon grid and 1752 timesteps):

variable	cdms2 deflated	CMOR3 no-deflation	CMOR3 level=1	CMOR3 level=3	CMOR3 level=9
areacello	32.8KB	303.7KB	41.4KB	-	-
siconc	29.3MB	433.1MB	44.6MB	43.2MB	39.9MB
siconcbcs	75.2MB	433.1MB	73.0MB	72.3MB	69.6MB
tos	225.9MB	433.1MB	177.1MB	174.8MB	171.3MB
tosbcs	239.2MB	433.1MB	185.1MB	182.6MB	178.4MB

It would be useful to have associated timing/CPU load equivalents to validate the statement by Balaji

[dnadeau4]

create test Test/cmor_speed_and_compression_03.py