PnetCDF buffers

[Rob Latham]

On 11/11/2015 10:03 AM, Max R. Dechantsreiter wrote:
> Dear PnetCDF:
>
> Based on experiments it appears PnetCDF establishes for
> each MPI process a file presumably used as a buffer -
> is that correct?

The file isn't what I'd call a buffer. It's the dataset holding whatever 
you wish to stuff inside it.

It's possible with some work to create one file per node, but it is more 
common to create a single file shared across all MPI processes via some 
kind of parallel file system.

You can provide a buffer to Parallel-NetCDF that it can use for 
aggregating operations, or you can use the non-blocking interface to 
queue up operations, in a sense buffering the operations in client-side 
memory.

==rob

>
> Best regards,
>
> Maxwhatev
> ---
> Max R. Dechantsreiter
> President
> Performance Jones L.L.C.
> max at performancejones.com
> Skype: PerformanceJones (UTC-06:00)
> +1 414 446-3100 (telephone/voicemail)
> http://www.linkedin.com/in/benchmarking
>

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA