topic Re: Altix-4700 and PBSPro thread migration errors in Intel® MPI Library

Altix-4700 and PBSPro thread migration errors

postaquestion — Wed, 09 May 2007 23:56:27 GMT

Hi - does anyone have a solution to this problem on an Altix-4700?

I am compiling my code with the Intel compiler and submitting to Altix-4700 using PBSPro. The job starts as per the PBSPro setup, i.e., on specified blades and on CPUs belonging to these blades. But, soon thereafter, the threads migrate to another set of processors. As a result, about 100 threads and system jobs are running on 44 cores, i.e., 22 Itanium 2 processors, leaving behind about 80-86% CPUs idle.

Can you please suggest what I can do to fix?

Thanks.

Re: Altix-4700 and PBSPro thread migration errors

TimP — Thu, 10 May 2007 14:10:38 GMT

This would depend on some issues which you haven't defined, such as which MPI you use, and on Altix specific options, which few of us know in detail. If you are using Intel MPI, youcould considersubmitting a fuller description of your issue on premier.intel.com under Intel MPI.

Re: Altix-4700 and PBSPro thread migration errors

dkokron — Fri, 11 May 2007 22:11:28 GMT

You left out some needed information. Is this an MPI application? What MPI library are you using? If you are using SGI MPI, then setting MPI_DSM_DISTRIBUTE will pin MPI processes to cores. If it's a hybrid code (MPI+OMP), make sure you also set the MPI_OPENMP_INTEROP environment variable. Use 'man mpi' for more details.

Dan