[Beowulf] SC|05: will 10Ge beat IB?

[Mark Hahn]

> >24pt 10G can be done on a single switch now, which necessarily 
> >brings its price down - there were a couple examples of this at 
> >SC05.
> 
> Who did these? I didn't get to the whole floor so it's likely I missed
> them.

fulcrum (which was not very handily located) was one.  I wasn't
trying to get specifics or prices, simply because we've already
committed our money...

> >Myricom offers something similar.
> 
> However, their switch is mx based. I never ssaw a 10 GigE switch
> from them (but I never got to spend much time with them).

Myri definitely talks about mixing eth and mx in the same switch;
I don't know the details or schedule, though.

> >>Another company is developing a new 10 GigE switch ASIC.
> >>Fulcrum Microsystems is developing a new ASIC with great
> >>performance (200 ns latency) and 24 ports. The general cost
> >
> >I asked about this a bit - the latency sounds great, but if you 
> >expect a 10G nic to cost ~10 us, it hardly matters whether the 
> >switch is 200 or 600 ns.
> 
> Unless you are doing fat trees with 3 or more tiers.

why?  1.8 us vs .6 us seems pretty easy to ignore if the nic 
costs you 10 us in the first place.

> It's also important for storage IMHO.

really?  I've heard that latency is important for lock manager
traffic, but it is certainly not for bulk transfers.  and for locks,
I imagine this is largely a question of the LM's design (leases,
hierarchical locks), rather than inherent to all LM's.