[mpich-discuss] process/core binding

[Pavan Balaji]

Hi Mark,

You can either use a user-defined mapping or Hydra's CPU topology 
functionality.

See 
http://wiki.mcs.anl.gov/mpich2/index.php/Using_the_Hydra_Process_Manager#Process-core_Binding

The following options should work:

/* Use appropriate values for x, y */
% mpiexec -np 2 -binding user:<x,y> ./foo

or

% mpiexec -np 2 -binding cpu:sockets ./foo

Note, however, that there were some bugs in mpich2-1.3.1 in this part of 
the code which got fixed after the release. So, if things don't appear 
to work correctly, I'd recommend using the latest nightly snapshot.

http://www.mcs.anl.gov/research/projects/mpich2/downloads/tarballs/nightly/hydra

These fixes should be available in the upcoming 1.3.2 release as well.

  -- Pavan

On 01/13/2011 08:00 AM, Mark Beauharnois wrote:
> Hi,
>
> Seems like my previous email was truncated after the table so I’ll repost…
>
> I have a 2x Quad core Intel Nehalem server, and /proc/cpuinfo shows the
> following processing element mapping (16 virtual processing elements I
> guess due to Intel’s Hyperthreading architecture):
>
> Physical ID 0 Physical ID 1
>
> Proc Id/Core Id Proc Id/Core Id
>
> 0/0 4/0
>
> 1/1 &n sp; 5/1
>
> 2/9 6/9
>
> 3/10 7/10
>
> 8/0 12/0
>
> 9/1 13/1
>
> 10/9 &nb sp; o:p>
>
> 11/10 15/10
>
> I would like to run 2 processes on the machine, and insure that each
> physical CPU runs one process. By default, does Hydra distribute the two
> processes over the physical CPU’s, or would it run the two processes on
> processor ID’s 0, and 1 (2 cores on same physical processor)? If that is
> true, how would I specify a user defined process/core binding in this
> situation to make sure that the processes are distributed across
> physical CPU’s? Would I use a binding such as “binding=user:0,4”?
>
> Thanks,
>
> Mark
>
> Mark Beauharnois
>
> Atmospheric Sciences Research Center
>
> State University of New York at Albany
>
>
>
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji