[Beowulf] Intel CSA performance?

Robert Myers rmyers1400 at comcast.net
Fri Mar 19 22:58:53 EST 2004

Greg Lindahl wrote:

>On Thu, Mar 18, 2004 at 10:26:51AM -0500, Mark Hahn wrote:
>>Intel added a special connection on their chipset to connect 
>>gigabit on some chipsets (CSA).  I've been wondering whether 
>>this would offer a latency advantage, since it's conventional wisdom
>>that PCI latency is a noticable part of MPI latency.
>Eh? PCI latency can be noticable when you have a low latency network,
>but gigE latency isn't nearly that low, especially once you've gone
>through a switch.
>The only reference to gigabit latency in the article didn't say what
>they measured. I'd assume that it was using the normal drivers, which
>means the kernel networking stack, which means you're looking through
>the telescope from the wrong end.
I had thought it might be interesting to fool around with trying to use 
CSA for hyperscsi, but I think you're saying if you're going to use a 
switched network, don't bother, if you're trying to win on latency.

When Intel abandoned infiniband and the memory controller hub sprouted 
this ethernet link, I figured that was their opening shot in stomping 
what's left of infiniband. Maybe it is, and they just don't care about 
latency, but it sounds like nobody's got any reliable information as to 
what the latency effects of CSA may be, anyway.

Every indication I can find is that Intel has all its bets on ethernet, 
and I don't know that there is any technological obstacle to building a 
low-latency ethernet.


Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

More information about the Beowulf mailing list