I'll let the Myricom guys answer further for message passing
optimizations, but some general tips:
- Yes, using processor affinity might help. Add "--mca
mpi_paffinity_alone 1" to your mpirun command line and see if that
- I *believe* that HP MPI uses processor affinity by default -- Open
MPI does not.
- If you are using less MPI processes than you have processors, you
might want to spread out the MPI processes across computational
sockets. Open MPI 1.3.x does not yet have good controls for this yet;
you might need to do this manually with some shell scripting (we're
putting better processor affinity controls in future versions).
On Aug 27, 2009, at 11:13 PM, <Bastil2001_at_[hidden]>
> Dear openmpi-ers,
> I lately installed openmpi to run OpenFOAM 1.5 on our myrinet
> cluster. I
> saw great performence improvements compared to openmpi 1.2.6,
> however it
> is still little behind the commerical HPMPI.
> Are there further tipps for fine-tuning the parameters to run mpirun
> with for this special application? From my experience the MX-ML should
> be the quicker one so I currently use:
> mpirun --mca mtl mx --mca pml cm ...
> as given the FAQ.
> I also thing that processor affinity might be worth trying, I will do
> this. Some other tipps? Are there special reasons why HPMPI still
> outperforms openMPI for this kind of tasks? Thanks and regards.
> users mailing list