Using GPU-Direct with pnetcdf?

[jedwards4b]

Has anyone tried this?

[wkliao]

Hi, Jim
I think it would be a great feature in PnetCDF.
Could you point me to some use cases that can benefit from this feature?

[jedwards4b]

I am working on an asyncronous implementation of the parallelio library in cesm. I would like to prototype parallelio moving data from cpu memory to gpu memory, doing some data compression and/or data postprocessing on the gpu and then writing directly from the gpu to disk.

[wkliao]

Can I assume the parallel write part of your use case is the typical concurrent writes from multiple processes to a single shared file? e.g. using collective I/O mode.

Just a reminder, PnetCDF (and the classic NetCDF file formats) does not support data compression.