[mvapich-discuss] Re: mvapich2 and Intel MPI Benchmarks

[Bryan Putnam]

On Fri, 18 Sep 2009, Bryan Putnam wrote:

> Hi All,
> 
> I just wondered if you'd had a chance to try out the Intel MPI Benchmarks 
> with mvapich2. IMB-3.2 is what I'm currently running.
> 
> We've had no problems with the latest mpich2-1.1.1p1, however when using 
> mvapich2-1.4rc2 or mvapich2-1.4rc1 (on both our IB and iWARP clusters) 
> the benchmarks are failing, for example even a simple
> 
> mpiexec -np 8 ./IMB-MPI1
> 
> will result in errors such as

I've discovered that these errors don't occur if each of the 8 processors 
is on a separate node. In fact if I disable "shared memory collectives" 
by setting the variable

MV2_USE_SHMEM_COLL=0

then the Intel MPI Benchmark seems to be happy even when the processors 
are on a single node, and things work fine on both the IB and iWARP 
clusters.

Bryan

> 
> 
> rank 6 in job 1  coates-a279.rcac.purdue.edu_51100   caused collective 
> abort of all ranks
>   exit status of rank 6: killed by signal 9 
> rank 5 in job 1  coates-a279.rcac.purdue.edu_51100   caused collective 
> abort of all ranks
>   exit status of rank 5: killed by signal 11 
> rank 4 in job 1  coates-a279.rcac.purdue.edu_51100   caused collective 
> abort of all ranks
>   exit status of rank 4: killed by signal 11 
> rank 3 in job 1  coates-a279.rcac.purdue.edu_51100   caused collective 
> abort of all ranks
>   exit status of rank 3: killed by signal 11 
> rank 2 in job 1  coates-a279.rcac.purdue.edu_51100   caused collective 
> abort of all ranks
>   exit status of rank 2: kille
> 
> 
> 
> Thanks,
> Bryan
> 
> --
> Bryan Putnam
> Rosen Center for Advanced Computing, Purdue University
> Young Hall (Rm. 519)
> 302 Wood Street
> West Lafayette, IN 47907-2108
> Ph 765-496-8225 Fax 765-494-0566
> bfp at purdue.edu
> http://www.rcac.purdue.edu
> 
> 
>