[Lustre-discuss] Parallel HDF5

[Andreas Dilger]

On 2012-07-18, at 3:24 PM, Tiago Soares wrote:
Dear all,
> I would like to know if there is some trick to use parallel MPI/IO in HDF5 file on Lustre?
> I have been trying for while to fix the issue that happens in Lustre. 
> 
> Basically, I have 8 process writing small datas in the same dataset in an attempt.  Each attempt write different data than before, and when an attempt comes close to 2500, the I/O parallel process broken. There is too a another serial process writing other data in the same file
> 
> 
> I found that is a common Lustre issue which do not support locking! So I tried this parameters above, but my application still broken.

Sorry, no "parameters above", but to enable distributed POSIX locking you need to mount clients with "-o flock".

>  Also, I read in somewhere to set the "stripesize" to count 8 (number of OST) for parallel IO, but still doesn't woks.
>  MPI_Info_create(&info);
>  
>  /* Disables ROMIO's data-sieving */
>  MPI_Info_set(info, "romio_ds_read", "disable");
>  MPI_Info_set(info, "romio_ds_write", "disable");
> 
>  
>  /* Enable ROMIO's collective buffering */
>  MPI_Info_set(info, "romio_cb_read", "enable"); 
>  
> MPI_Info_set(info, "romio_cb_write", "enable");
>   https://wickie.hlrs.de/platforms/index.php/MPI-IO 
> 
> I ran my application on PVFS, and works fine. I know that unlike the Lustre,  PVFS use non-alligned striping data. Could be it the reason that works in PVFS?
> 
>  Regards
> 
> -- 
> Tiago Steinmetz Soares
> MSc Student of Computer Science - UFSC
> 
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss


Cheers, Andreas
--
Andreas Dilger                       Whamcloud, Inc.
Principal Lustre Engineer            http://www.whamcloud.com/