Myrinet vs FastEthernet

[Greg Lindahl]

> For MM5, it's disappointing : 15% one proc per node, 10% with 2
> procs per node. The runtime (RSL) seems to be expensive (on top of
> MPI). The code itself seems to try to overlapp, so the SMP mode is
> more expensive that the single CPU mode. Greg Lindhal uses Myrinet
> for this code if i remember well, for scalability issues i guess
> (Greg ?).

Yes, scalability. MM5 scales nicely to 128 processors on fairly small
problems with Myrinet. Typical production MM5 runs on the FSL machines use
36 or more cpus at a fairly small problem size, and will use more as we add
more processors. And these are Alphas, so if you wanted to compare to x86,
think 72+.

Since you only had 8x2 cpus, you didn't get far enough out to see much
benefit from a better network. It would be nice to see your paper done with
more processors, but even with just 8x2, it sounds like very good work. And
anyone who wants to run MM5 on a small machine now knows...

-- greg