MPI performance on Coral
josip at icase.edu
Wed Oct 11 15:02:11 PDT 2000
I have assembled the MPI performance results obtained on our cluster
We tested MPI/Pro, MPICH+MVICH and LAM over Fast Ethernet (TCP), local
(loopback or shared memory) and Giganet cLAN (VIA) transports. All of
the tested MPI libraries reach high bandwidth; the differences are
primarily in latency and in CPU cycles remaining to the application on
dual CPU nodes.
While collective communication tests with MVICH have not been completed
yet, I hope to add them shortly. The rest of our data is already on
line: our best result was 12 microsecond MPI latency and 107.2 Mbyte/s
bandwidth using MVICH over Giganet cLAN.
P.S. Earlier, we also tested Gigabit Ethernet using 3com/Alteon/NetGear
cards, but found that reasonably high performance could only be reached
under very contrived circumstances. Fast Ethernet bonding should do
about as well at a much lower cost. At a somewhat higher cost, Giganet
or Myrinet should perform much better.
Dr. Josip Loncaric, Senior Staff Scientist mailto:josip at icase.edu
ICASE, Mail Stop 132C PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134
More information about the Beowulf