[mpich-discuss] mpirun on 1500~2000 cores

[dvg]

> significantly improves the time taken by MPI_Init

Much better, 40x improvement, from 20 minutes into ~30 seconds.
(I tried mpich2-trunk-r4899.tar.gz)

Thanks,
Dmitry.


On Sun, 2009-07-05 at 00:24 -0400, Rajeev Thakur wrote: 
> There was a fix applied recently that significantly improves the time taken by MPI_Init on large numbers of processes when using the
> Nemesis communication channel and the MPD process manager (both are the default options). The fix will be included in the 1.1.1
> release due out late next week.
> 
> In the meanwhile, please try out one of the nightly snapshots of the svn source from
> www.mcs.anl.gov/mpich2/downloads/tarballs/nightly/trunk/ and let us know if it improves the time taken to start your job.
> 
> Thanks,
> Rajeev
> 
> 
>  
> 
> > -----Original Message-----
> > From: mpich-discuss-bounces at mcs.anl.gov 
> > [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of dvg
> > Sent: Saturday, July 04, 2009 10:04 PM
> > To: mpich-discuss at mcs.anl.gov
> > Subject: [mpich-discuss] mpirun on 1500~2000 cores
> > 
> > Hello,
> > 
> > What would be considered as reasonable time for mpirun to 
> > start a job on 1500~2000 cores, 1 gige cluster?
> > 
> > Are there any kernel (linux) or eth-related parameters which 
> > can be tuned to speed it up?  MPICH2 libraries were compiled 
> > with most/all optimization options enabled.
> > 
> > Thank you,
> > Dmitry
> > 
> > 
> 
>