Lustre discuss - Dec 2007 - MPI-Blast + Lustre

Anyone have any experience with MpiBlast and Lustre.     We have  
MpiBlast-1.4.0-pio and lustre-1.6.3 and we are seeing some pretty  
poor performance with most of the mpiblast threads spending 20% to  
50% of their time in disk wait.    We have the genbank nt database  
split into 24 fragments (one for each of our OSTs, 3 per OSS).    The  
individual fragments are not striped due to the ldlm_poold issue so  
that should not be the problem.

Should we not be using the PIO (ROM-IO) version of MPIBlast w/ lustre?

I''m wondering if there are some already discovered guidelines for  
using MPIBlast on top of Lustre that I''m not finding via google.

Thanks,

Charlie Taylor
UF HPC Center

Hello,

Charles Taylor wrote:
> Anyone have any experience with MpiBlast and Lustre.     We have  
> MpiBlast-1.4.0-pio and lustre-1.6.3 and we are seeing some pretty  
> poor performance with most of the mpiblast threads spending 20% to  
> 50% of their time in disk wait.    We have the genbank nt database  
> split into 24 fragments (one for each of our OSTs, 3 per OSS).    The  
> individual fragments are not striped due to the ldlm_poold issue so  
> that should not be the problem.
>
> Should we not be using the PIO (ROM-IO) version of MPIBlast w/ lustre?
>   
I am not familiar with MpiBlast implementation. Does it use collective I/O ?
Could you please specify its I/O pattern here?
And also it is always worth trying to adjust these hints, if  you use  
MPI  I/O.

disable/enable romio_cb_write 
disable/enable romio_ds_write
ind_wr_buffer_size and ind_rd_buffer_size

Thanks
WangDi
> I''m wondering if there are some already discovered guidelines for
> using MPIBlast on top of Lustre that I''m not finding via google.
>
> Thanks,
>
> Charlie Taylor
> UF HPC Center
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at clusterfs.com
> https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
>