[Beowulf] 1.2 us IB latency?

[Mark Hahn]

> Nothing was broken in the previous InfiniBand adapters. The previous
> generation, with higher MPI latency still beat other solutions that
> shows lower latency, due to the fully offload architecture.

look, that's just not true.  I've got a cheap, low-end cluster which 
uses plain old myri2g and mx, and has for ~3 years.  3 us latency.
IB was more like 6-8 us (at the mpi level, of course) three years ago,
and with the exception of your new promised adapters, is still not 
faster than 2g-mx...

> We gain experience from each generation and implement it in the next
> generations, and this is the outcome.

thanks, I needed my daily dose of marketing-speak.  unfortunately,
the question remains unanswered.  gradual improvement does not explain 
a 3x improvement.

> There are no special requirements for achieving this MPI latency, and
> we are very happy to provide low latency without changing the offload
> concept of our architecture.

OK, so what's unique about your offload concept?  it's obviously not the case
that you're the first to do offload.

>> also, just to be perfectly explicit, this is 1.2 us
>> inter-node, right?  not something crazy like two 8-core boxes
>> with only two of 16 hops inter-box?
>
> 2 nodes connected via InfiniBand - regular setting. There is no
> dependency on the
> numbers of cores, as we don't need to CPU to drive the interconnect.

sniping at infinipath aside, thanks for the confirmation.