[mpich-discuss] Cluster with uneven CPU speeds

[Jed Brown]

"Balaji, Pavan" <balaji at anl.gov> writes:
> If you want to try the approach with oversubscription, I’d recommend
> configuring mpich using --with-device=ch3:sock.  That build will not
> be as optimized for the common case of dedicated cores for each MPI
> process, but also won't hurt the oversubscription case as much.  I
> already warned you that this is a bad idea, so use at your own risk.
> :-)

It's a bad idea for performance, but great for testing and debugging.
It's crazy to not be able to run a 9-process correctness test on a
laptop.  For this reason, PETSc's --download-mpich feature (used by
people that don't have MPI installed, or whose OS provides a broken MPI)
still uses ch3:sock by default.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20140606/1b2cc806/attachment.sig>