Choice of Zarr compressor

[VeckoTheGecko]

The writing of the zarr file in particlefile.py doesn't look to set a compressor. Setting a compressor can significantly help trading off compute for storage or vice versa [1]. The default chosen by zarr is the Blosc compressor, which is a "meta-compressor" (using different algorithms under the hood). Perhaps its worth investigating other compressors to see if there's one that is best suited for our simulation output.

@erikvansebille do you know if zarr compressors have been a topic of discussion before?

[JamiePringle]

All the Zarr files I have made from parcels have been compressed. You need not set a compressor; the default is to compress. For efficiency I have found the correct setting of chunk size to be more important.

[VeckoTheGecko]

I mean the type of compression that is used (there are various available compressors zarr-developers/numcodecs). I agree about chunk size, but it would be good to also investigate the compression algorithm used as I think we just rely on the default which may not be best given our data.

[CKehl]

With all due respect, I get the impression you miss the point of the paper linked above. The high I/O-time is attributed in major parts to the loading of the fieldsets. Correct if I am mistaken, but that doesn't have anything to do with the zarray compressor of the particle file, does it ? I wish you get the answers you seek through your benchmarking approach.

[VeckoTheGecko]

Thanks for clarifying! :)

I haven't had the time yet to fully go through the paper. I've updated the description here to match.

[VeckoTheGecko]

Closing this. In v4 of Parcels users can choose their compressor for output data using https://zarr.readthedocs.io/en/stable/user-guide/config.html#runtime-configuration . If a user finds a better compressor than the default for some reason or another - they can open an issue (where we can then update a tutorial recommending to set a certain compressor).