[Beowulf] RDMA NICs and future beowulfs

[Greg Lindahl]

On Mon, Apr 25, 2005 at 08:26:33PM -0400, Joe Landman wrote:

> Not sure I agree with this.  If there were no value in TCP offload, then 
> why would Intel announce (recently) that they want to include this 
> technology in their future chipsets?

If you go read the research papers of the guy who invented I/OAT, he is
adding minimal hardware features to get nearly the same benefit as
offload, without offloading. His approach is similar to PathScale's
approach. Intel's marketing is to position I/OAT between dumb nics
and TCP offload engines (TOEs). TOEs are much more expensive and
Intel's marketing position is that they're going to be a niche.

Now I'm not going to debate the merits of Intel's marketing position,
but I wanted to point out that it didn't match at all what you were
saying.

> Most of the codes we play with are latency bound.

Then you should note that offload generally doesn't improve latency.
Offload's main strength is high bandwidth with less overhead. There
are several examples of Linux software that gives low latency on
standard ethernet hardware: Scali, Par-Tec, Gamma, MVIA.

-- greg