[mpich-discuss] mpiexec.hydra - process binding limitation?

[Pavan Balaji]

Currently, Hydra doesn't have the ability to bind to multiple cores (we 
had this earlier, but disabled it recently).  We will be revamping our 
binding model in the 1.5 release to cover these cases.

  -- Pavan

On 05/15/2012 03:47 AM, Anatoly G wrote:
> Dear MPICH2.
> I use Intel processor with NUMA technology.
> All processes are multi-threaded.
> Processes distributed between computers by mpiexec.hydra.
> I"m interested to separate processes by memory usage. Each process use
> it's local memory only, as result bound to single socket (but not core).
> But I"m interested in leaving maximum freedom to OS in choosing cores to
> achieve maximum performance. As result, I"m interested in binding to
> CPU:socket but not core.
>
> For example:
> I have computer with 2 CPUs,  each CPU has 4 cores.
> *) I try to execute 2 processes on this computer.
>      I"d like to bind 1-st process to cores 0-3 (OS decides where it
> will be executed each time-slot) and
>      2-on one on cores 4-7.
> *) I try to execute 4 processes on this computer.
>      Processes 1&2 I want to bind to CPU:Socket 1 (cores 0-3), processes
> 3&4 I want to bind to
>      CPU:Socket 2 (cores 4-7). OS will resolve conflicts.
>
> According to manual, each strategy performs binding to specific single
> core, which calculated using user selected algorithm.
> 1) May you please suggest any technique to achieve expected binding (not
> to single core), using mpiexec.hydra?
> 2) Can I perform process binding to local memory per CPU:Socket?
>
> Thank you,
> Anatoly.
>
>
>
>
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>

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