>>>>> "Paul" =3D=3D Paul M Newhouse <newhouse@rockhead.com> writes:

    Paul> Chris Jones writes:
    >> My actual question is, can we do something like this with
    >> NetBSD?  Really, we just want to maximize the bandwidth/dollar
    >> ratio.  ATM looks expensive, and 100bT may not have the
    >> bandwidth we need.  Is there a way to multiply your bandwidth?
    >> 
    >> Any thoughts?

    Paul> Look into HiPPI (esh driver) I think it's checked in?
    Paul> 100MBytes/s capability.  I think the Pentium II's on P6DNH
    Paul> boards are getting about 30MBytes/second.  It's a bit
    Paul> pricier but, you get more performance with it.  Look at:

    Paul> GigaLabs - www.gigalabs.com

    Paul> Essential(Optical Data Solutions) - www.ods.com

I don't agree with this. 

We're running HiPPI among our large beasts, but we're still using
FastEthernets for the Beowulf-PC-cluster-thingy we're currently
building. The whole point with PC clusters is the cost benefits of
cheap hardware. HiPPI is now only a fraction of the cost it was a few
years ago, and the bandwidth/dollar ratio isn't all bad but it is
still simply too expensive to use as interconnect in a sizeable PC
cluster.

And also remember that although the ratio is nice if you're using the
800Mbit/s theoretical bandwidth you won't see anything near that on a
PC. On the large SGIs, though, we get much closer, but of course they
cost much more...

Our PC cluster is (presently) running Linux (sorry) and one major
reason for this is the channel bonding stuff that NASA did. So, I
definitely agree that loadbalancing over multiple cheap interfaces is
an interesting thing to have and I'd love to see it in NetBSD.

Regards,

Johan Ihr=E9n, <johani@pdc.kth.se>, Center for Parallel Computers,
Royal Institute of Technology, S-100 44 Stockholm, Sweden