[mpich-discuss] Scalability of Intel quad core (Harpertown) cluster

Hee Il Kim heeilkim at gmail.com
Wed Apr 23 23:45:01 CDT 2008

Thanks All,

Only recently I could have a time to change the configuration of switch hub
and test this and that. Unfortunately, I failed to take the advantage of
Open-MX, which could not run Cactus benchmark for some reasons.

But after changing the maximum frame size to 9000 and enabling flow-control,
the latency could be reduced down greatly,

## mpi_pong reuslts

  262144 bytes took       708 usec ( 740.412 MB/sec)
  524288 bytes took      1500 usec ( 699.101 MB/sec)
 1048576 bytes took      3523 usec ( 595.256 MB/sec)

 Max rate = 1139.390288 MB/sec  Min latency = 2.980232 usec


And also I could get reasonable scalability for Cactus benchmarks. The MTU
of the ethernet cards should be taken 1500, the default one. If I increase
the value, it produces worse results.

In conclusion, the culprit of the scalability problem was the interconnect I
used, not the cpus or softwares. I appreciate all your comments and


Hee Il
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20080424/f01242eb/attachment.htm>

More information about the mpich-discuss mailing list