New class of Beowulf clusters ? (Re: two NICs - channel bonding -

[Greg Lindahl]

> Folks in CSC, Helsinki, Finland have measured 312 barriers/sec for 32
> processes running on 16 dual PentiumsII 400Mhz connected by switched
> fast ethernet (under MPI over TCP/IP). It clearly limits usability of
> such a cluster only to (very) coarse grain parallel programs.

Tweet! The overgeneralization police have given you a ticket.

PAPERS does one thing really well: global barriers and broadcasts.
That's it. The benchmarks I've been doing all involve algorithms which
would not be sped up by PAPERS. Most of them involve either nearest
neighbor communication, or all-to-all, where each process receives
different information.

It would be nice if someone who did have an algorithm which is helped
by PAPERS integrated it with MPICH. That would make it easier for the
rest of us to experiment with it. But it isn't a magic bullet.

> ** Does someone know, how many barriers can by done over MYRINET or **
> ** using U-Net ?                                                    **  

Probably about only 5x as many possible with TCP/IP. However, you
could hack some features into your myrinet driver to directly support
broadcasts, and that would probably speed it up by quite a bit. The
fastest possible would probably be around 10,000 per second, which is
far slower than PAPERS.

-- g