[Beowulf] 1.2 us IB latency?

[greg.lindahl at qlogic.com]

On Thu, Apr 19, 2007 at 10:35:56AM +0200, Håkon Bugge wrote:

> This benchmark was _interesting_ in the sense it 
> showed Gbe over the TCP/IP stack to have lower 
> CPU overhead than Infiniband...

I'm not sure exactly which results you are referring to, but keep in
mind that you need to divide the overhead by the number of bytes
sent. GigE usually sends a lot fewer bytes than InfiniBand.

> Obviously, this because it uses only one MPI process per node - and
> in my opinion is flawed by that.

I agree that using all the cores is the right approach; that's what
real programs do. InfiniPath is embarrassingly parallel in that
dimension.

-- greg