[Beowulf] Re: torus versus (fat) tree topologies
hahn at physics.mcmaster.ca
Mon Nov 22 11:22:37 PST 2004
> Which is one of the nice things about InfiniBand, industry standard;
> more vendors more competition.
IMO, "industry standard" doesn't apply until there are at least
three independent sources of all necesary parts, and really it's
a bit of a lie until random far-eastern factories can churn out
so if Mellanox disappears from the face of the world tomorrow,
I think there would be much gnashing of teeth in the IB world,
which to me means that the standardness is entirely de-facto.
ie, only marginally more standard than Myrinet or even Quadrics.
> We should also begin to see multiple RNIC (RDMA over Ethernet) vendors
> in the HPTC interconnect market very soon.
with what latency? as far as I can tell, people in that market are
under the delusion that TOE and RDMA is all matters. sure, burning
fewer host cycles is good, but it doesn't make eth compete with quadrics.
and RDMA hardware is nice, but unless it drastically improves latency,
it's basically silly (ie, existing TCP SACK is not far from being RDMA,
and it's a SMOP to do RDMA using a conventional non-RDMA nic.)
More information about the Beowulf