[mvapich-discuss] MVAPICH2 1.5 hwloc issue

[Dhabaleswar Panda]

Hi Eoin,

Thanks for your detailed note. We are working on MVAPICH2 1.5.1 release
today. This release has a fix to take care of CPU mapping with HWLOC
support and SMT. I am hoping that this will resolve the issue you are
seeing. It will be good if you can try the release once it is made and let
us know whether it solves your problem or not.

Best Regards,

DK

On Tue, 14 Sep 2010, Eoin McHugh wrote:

> On Tue, Sep 14, 2010 at 01:43:56PM -0400, Jonathan Perkins wrote:
> > Thanks the for the note.  We've recently resolved an issue where our
> > internal version of hwloc was not being properly built in our shared
> > library.  Can you describe your issue a bit more.  Are you seeing
> > build or linking failures or unexpected operation at run time?
>
> Hi Jonathan,
>
> Thanks for your reply. I'm seeing unexpected operation at runtime. I've
> a dual socket hex-core Westmere system which happens to have hyper
> threading  enabled. I'm building HPCC against MVAPICH2 and observing it's
> behaviour with 12 processes on a compute node. With either the internal
> hwloc and an external static library version of hwloc I'm seeing random
> non-deterministic process placement. With an external shared library
> hwloc processor affinity works and can be controlled with the usual
> MV2_* variables.
>
> I'm attempting to build the default static library only version of
> MVAPICH2. I tried a shared library version and the behaviour was the
> same as above.
>
> Regards,
>
> --
> Eoin McHugh
> ICHEC Systems Team
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>