[Beowulf] 1.2 us IB latency?

[Mark Hahn]

>> a dramatic breakthrough: 1.2 us, 25M msg/s.  since we just happend to
>>
>> or is this an example of message aggregation?  heck, from the url
>> above, it might even be counting intra-box messages.
>
> Nope, this is "normal" ping-pong for the new generation cards (connectx).

so if I do this:

start timer 
send(other,small-message)		recv(first,small-message)
recv(other,small-message)		send(first,small-message)
stop timer

I'll actually see 2.4 us between the timer calls?  if I understand,
aggregation would only help on a streaming test.  in fact, this kind 
of isolated RPC-like exchange is what I see most commonly.

> Maybe a bit optimistic though, I'd expect closer to 1.5 in a back-to-back
> config.

so for a small switch (24pt, say), how many hops to the internal fabric,
and they're, what, .2 us each?

also, does back-to-back work well?  I can imagine some cases where 
putting two dual-port cards in each node and creating a mesh might 
work well.

>> also, I'm sorta amazed people keep selling (and presumably buying)
>> dual-port IB cards.  doesn't that get quite expensive, switch-wise?
>
> Not defending them but, It could possibly maybe be useful if you have a
> stand-alone IB net for, say, storage or something else not mpi. Also, it's
> not like they're that much more expensive than single port ones...

yeah, I can see PHB's buying redundant fabrics.  I'd be more interested in
using the higher port-count for FNN or related topologies (assuming switches 
are cheap, at least at some size...)