For the web archives...
Brock and I talked about this in person at SC. The conversation was much more involved than this seemingly-simple question implied. :-)
The short version is:
- numactl does both memory and processor binding
- hwloc is the new numactl :-)
- e.g., see the hwloc-bind(1) command
- OMPI does both memory and processor binding
- OMPI 1.5.5 will have an MCA parameter for process-wide memory binding policy
- Torque cpusets are probably do what is desired: restrict MPI processes to a subset of the processors on a given server (e.g., if multiple Torque jobs are running on the same server)
On Nov 9, 2011, at 1:46 PM, Brock Palen wrote:
> If we are using torque with TM with cpusets enabled for pinning should we not enable numactl? Would they conflict with each other?
> Brock Palen
> Center for Advanced Computing
> users mailing list
For corporate legal information go to: