In MPICH, OFI vs. UCX makes no difference for shared memory since that’s a different module. CH4 vs. CH3 is the only choice that matters in that case.
CH3 vs. CH4 shared memory is pretty close, but it most instances, CH3 is still a little faster. There’s some changes in progress to improve things, but they’re not ready yet.
Thanks,
Wesley
> On Dec 14, 2019, at 6:51 PM, Jeff Hammond via devel <devel(a)mpich.org> wrote:
>
> MPICH configure just asked me to choose ch4:ofi vs ch4:ucx vs ch3. Does anyone have an informed opinion on which one is the faster for shared-memory execution?
>
> My specific use case is NWChem on very large multi-socket Xeon nodes, where passive target RMA is the primary communication method, hence my primary concern is asynchronous progress and lack of serialization in RMA. In the past, I have observed significant performance issues due to serialization of RMA accumulate operations acting on non-overlapping memory regions.
>
> Jeff
>
> --
> Jeff Hammond
> jeff.science(a)gmail.com <mailto:[email protected]>
> http://jeffhammond.github.io/ <http://jeffhammond.github.io/>_______________________________________________
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/devel
MPICH configure just asked me to choose ch4:ofi vs ch4:ucx vs ch3. Does
anyone have an informed opinion on which one is the faster for
shared-memory execution?
My specific use case is NWChem on very large multi-socket Xeon nodes, where
passive target RMA is the primary communication method, hence my primary
concern is asynchronous progress and lack of serialization in RMA. In the
past, I have observed significant performance issues due to serialization
of RMA accumulate operations acting on non-overlapping memory regions.
Jeff
--
Jeff Hammond
jeff.science(a)gmail.com
http://jeffhammond.github.io/
