Gigabit Switch

Greg Lindahl lindahl at pathscale.com
Tue Nov 11 19:37:38 EST 2003


On Tue, Nov 11, 2003 at 04:47:50PM -0500, Mark Hahn wrote:

> TCP is good at dealing with out-of-order.  practically by by definition!

Oh? Please share some test results. The reality is that out-of-order
packets are a moderate load on the CPU at best, and Linux isn't
exactly great at handling them, especially with multiple cpus and
multiple interfaces.

> I think latency is the real appeal of hw-supported multicast - if you 
> want to do a barrier across 256 nodes, do you want a ~8-deep tree of
> user-level processes farming out your tinygrams (say, 8x50=400 us), 
> or do you want a single 30 us multicast?

A reliable barrier built using unreliable multicast isn't 30 usec. And
MPI programs rarely have barriers. Perhaps you know of an MPI
application that really needs a barrier operation? I don't think I've
run into one yet.

-- greg

_______________________________________________
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