[Beowulf] Correct networking solution for 16-core nodes

[Greg Lindahl]

On Thu, Aug 03, 2006 at 02:02:16PM -0700, Gilad Shainer wrote:

> There was a nice debate on message rate, how important is this
> factor when you Want to make a decision, what are the real
> application needs, and if this is just a marketing propaganda.

I would be happy to compare real application performance, if you'd
like to do that instead. You could start by running any of the
applications in the 2 whitepapers we have on our website. Insomuch as
I was able to find Mellanox IB numbers, for example, I could only find
SDR numbers, but I'd love to have DDR numbers in addition.

> For sure, the message rate numbers that are listed on Greg web site
> regarding other interconnects are wrong.

Please send us a correction; that's what we observed on the cards and
software that we own, but if you can get a better number, we'll post
it. Mind you, the mpi_multibw page I just cited doesn't have *any*
numbers except those for InfiniPath. You're welcome to correct that if
we're wrong, but in addition if you find any numbers that are wrong
elsewhere (like in the whitepapers, which do have Mellanox message
rate numbers), we'd be happy to fix them.

We'd prefer that someone like DK Panda publish 3rd party message rate
numbers. That's why we give the code out.

-- greg

(I'm out the door for a 2-week vacation, so I second Christian Bell
and Kevin Ball to carry the torch while I'm gone. And Kibo, wherever
he is. But not Serdar Argic.)