[Beowulf] using two separate networks for different data streams
hahn at physics.mcmaster.ca
Sun Jan 29 22:18:22 PST 2006
> > I tend to think of IB as mainly for high-bandwidth, since afaikt its
> > latency isn't even as good as myri 2g (mx, 3 us). to say nothing of
> > myri 10g, quadrics, numalink, sci.
> I know some people don't think of InfiniPath as being InfiniBand, but
> really, Mark, I didn't think you were one of them! ;-)
hmm, I had the impression that InfiniPath operated in either fast
InfiniPath-to-InfiniPath mode, or in slow generic IB mode. that is,
that in "fast" mode, it was mostly just recycling IB phy layers.
> In any case, yes, you are correct for Mellanox InfiniBand HCAs, but
> not for InfiniPath.
so if Mellanox IB has latency X, and InfiniPath has Y,
the expected latency between the two is around (X+Y)/2?
More information about the Beowulf