Hello,
I have noticed that simple things like MPI_Barrier
will raise my CPU usage to near 100% while waiting,
and read in past emails and the FAQ that this is due
to busy polling in Nemesis.
As far as I can tell, the one solution usually offered
to reduce this CPU usage
is reconfiguring with sockets instead of Nemesis.
I would also like to use the MPI_THREAD_MULTIPLE
feature, but I read that the socket channels
don't support it, and it would be annoying to
have to recompile MPICH to get one benefit or the other.
I was just wondering if recent advancements invalidated
some of the above and officially ask what the options are
for reducing the CPU usage of MPICH during barriers.
Thank you,