[Beowulf] InfiniBand channel bundling?

[Jason Riedy]

And Mark Hahn writes:
> indeed, very nice.  though also quite surprising - is it known that
> FDR is so terrible for latency?  seems astonishing to me.

FDR increases bandwidth by increasing the on-wire packet size and
not any physical BW increase.  Therefore, if you're sending packets
that are too small, latency will suffer simply because you're still
waiting on the same physical interface to digest a larger chunk.
(Ok, that's more the 'o' term in LogP, but good enough.)

I'm guessing that those of us in the small-packet regime assumed
this from the get-go, so never really bothered checking.  The nifty
thing is that the posted results don't show *too* much loss for
BW-tuned dual-rail QDR.  (Um, ok, LogP shows that, too, but I'm lazy.)