[Beowulf] Q: IB message rate & large core counts (per node)?
Shainer at mellanox.com
Tue Feb 23 15:08:48 PST 2010
> The benchmark that we created is not a coalescing benchmark.
> Coalescing produces a meaningless answer from the message rate
> benchmark. Real apps don't get much of a benefit from message
> coalescing, but (if they send smallish messages) they get a big
> benefit from a good non-coalesced message rate.
Agree, most benefit will come from non-coalesced message rate - and for
the broader audience - from the ability to send one MPI message within a
single network packet. Message coalescing is when you incorporate
multiple MPI messages in a single network packet.
> If you look back in the archives of this list, you can find me saying
> that. And some other people were involved in the discussion, too.
> Or, we can take your word for what happened, instead of looking at
> history. I know which method you prefer.
You can look at history or not, your decision. My claim (and more
important for the present) is that the so called "non-coalescing"
results published on latest InfiniPath based products are actually with
coalescing. The other universe you called it?
More information about the Beowulf