[mvapich-discuss] theoretical limit to pt2pt communication
Peter Steinbach
steinbach at scionics.de
Thu Mar 17 04:33:13 EDT 2016
Hi to all,
our cluster is having more and more MPI users. For this, we recently had
to profile and debug the performance of our mpi installation. The OSU
benchmark suite from here:
http://mvapich.cse.ohio-state.edu/benchmarks/
was very instrumental for this process. So thank you very much for
providing it.
As a performance engineer, I however wonder what the theoretical limit
of say osu_mbr_mw is? In other words, if I get some performance X from
osu_mbr_mw, when should/must I stop researching the MPI/IB documentation
where to tweak?
To be more concrete, on our local Qlogic QDR 40Gbps network
infrastructure, osu_mbw_mr levels off at 3342.28 MBps with mvapich 2.1.
Apparently, this is the same value that is reported here:
http://mvapich.cse.ohio-state.edu/performance/pt_to_pt/
If I run the same benchmark linking to openmpi 1.10.2, osu_mbw_mr levels
off at 3961.54 MBps.
As it is very hard to correlate both numbers to the expected 5GBps that
I expect from QDR 40Gbps, I am a bit lost.
I'd appreciate your feedback!
Peter
PS. It would be nice if the plots on e.g.
http://mvapich.cse.ohio-state.edu/performance/pt_to_pt/
would connect somehow to which command in the OSU benchmarks they
correspond.
More information about the mvapich-discuss
mailing list