[mpich-discuss] MPI_File_sync and MPI_File_set_size

[Rob Ross]

You can't flush from rank zero in the no_indep_rw case, have to flush  
from first aggregator instead... -- Rob

On Dec 9, 2009, at 9:55 AM, Rob Latham wrote:

> On Wed, Dec 09, 2009 at 01:35:58AM -0600, Wei-keng Liao wrote:
>> In ROMIO, when the hint romio_no_indep_rw is set to true,
>> only the IO aggregators open/read/write/close the file.
>>
>> In MPI_File_sync, ADIOI_TEST_DEFERRED() is called before
>> ADIO_Flush(), to check if the file is opened, which forces
>> all non-aggregators to open the file. I think it is not
>> necessary. A simple check for open status can fix this, eg.
>>  if ( (fh)->is_open ) ADIO_Flush(fh, &error_code);
>
> Good point.  Furthermore, in most of the drivers (except ad_nfs and
> ad_bgl) we flush from rank 0 and broadcast the result.  And even in
> ad_nfs and ad_bgl, if the file isn't opened then there's nothing to
> flush from the cache.
>
>> Also, the similar issue happens for MPI_File_set_size().
>> The current implementation lets aggregator 0 to call
>> ftruncate() and then broadcast the result to all processes.
>> It is also not necessary to call ADIOI_TEST_DEFERRED().
>> The fix is simply remove the call.
>
> again, the only tricky point is AD_NFS, but if the file hasn't been
> opened yet, then the client can't have cached a file size.
>
> ==rob
>
> -- 
> Rob Latham
> Mathematics and Computer Science Division
> Argonne National Lab, IL USA
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss