[mpich-discuss] Hydra process affinity

[Jeff Squyres]

On May 25, 2012, at 10:10 AM, Martin Cuma wrote:

> I have thought that -binding cache:l3 would do that, i.e. pick all the cores on the socket, as they share the L3. However, that is not the case:
> /uufs/chpc.utah.edu/sys/pkg/mpich2/1.5b1/bin/mpirun -binding cache:l3 -np 2 /uufs/chpc.utah.edu/sys/pkg/hwloc/1.4.2/bin/hwloc-bind --get
> 0x00000002
> 0x00000001

If I could hazard a guess, it looks like the -binding <foo> option in hydra effects the process placement policy of each process, but does not change its "binding width" (e.g., how many hwloc-esque PU's to bind the process to).

-- 
Jeff Squyres
jsquyres at cisco.com
For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/