[Beowulf] New HPCC results and the Myri viewpoint
patrick at myri.com
Thu Jul 21 12:11:15 PDT 2005
Keith D. Underwood wrote:
>>Yes, Keith noted it also, it's useful to evaluate the receive rate of
>>N-to-1 pattern. I meant that it's useless to optimize the send side in
> Except that many nodes to both an N-to-1 and a 1-to-N. i.e. they
> receive from N neighbors and send to those same N neighbors in each time
> step. That means that both the send and receive sides of the streaming
> bandwidth issue matter.
I have never seen a code with 1-to-N where N is big enough to fill up
the pipeline. It's this type of pattern (N-to-1 is much worse than
1-to-N) that kills scalability in large config and developpers know
that. Even in a classic all-to-all, most implementation uses a spanning
log tree, some even with a fan-out limit just to limit the damages.
More information about the Beowulf