Re: MPICH-GM + MKL trouble
Craig Tierney <ctierney <at> hpti.com>
2002-12-11 16:14:55 GMT
On Wed, Dec 11, 2002 at 04:17:37PM +0300, Birukov Kirill wrote:
> I've got a problem running HPL (High Performance Linpack), compiled
> with Intel MKL, under MPICH-GM. I am trying to run it on four 2-way
> Xeon nodes, starting 1 process per node. I let MKL to parallelize on
> each node (i.e. MKL creates 2 threads per node and both CPUs are
> loaded). This trick works with MPICH from Argonne, which utilizes P4
> device, but not MPICH-GM. Under MPICH-GM I've got some messages about
> signal 11 after program start and it hangs. Has someone experience
> with MPICH-GM and MKL?
I have run mpich-gm and mkl, but not with multiple threads per process.
Other tests I have tried using multiple threads per process were never
as fast as run running multiple processes (mkl or autoparallelization of
When I tried mkl 5.0, I found that Altas produced much faster results
than mkl when running HPL.
Have you turned hyperthreading off? If not, are you starting 4 processes
or threads per node? AT SC2002, Dell presented a paper in their booth
where running 4 processes on 2 proc box with HT on gave a boost of 10% or
> P.S. When I limiting MKL with 1 thread per process, and running 2
> processes per node, everything work fine. But I have information that