IB vs. Myrinet
ktpedre at sandia.gov
Tue Nov 4 12:35:40 EST 2003
> Subject: IB vs Myrinet
> Date: Mon, 3 Nov 2003 22:19:48 -0500
> From: "Joey Sims" <jsims at csiopen.com>
> I believe IB is a much better interconnect technology than Myrinet
> period. Plus, you don't have to deal with Myricom.
Things like Myrinet and Quadrics do have at least one architectural advantage
over IB for HPC -- they have a programmable processor on the NIC. Myricom's
MX will, presumably, use the NIC processor to offload MPI receive matching.
Quadrics Tports also offloads MPI matching. Offloading theoretically lowers
host CPU overhead (less interrupts) and lowers latency (less trips across the
PCI bus). If Ohio State's MVAPICH really scales beyond 8 nodes well (I've
only seen 8 node benchmarks), then maybe my point is irrelevant. Still, in
my opinion the offload approach is more elegant.
I've heard some IB HBAs are programmable but there is no standardization and
documentation is scarce. Does anybody have more information?
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