Re: [Myrinet] questions

On Tue, Jun 11, 2002 at 02:32:06PM -0400, Bill Allard wrote:

> Hmm... This is better than I thought.  Just to make sure, you seem to 
> be saying that <at the time> when GM_SUCCESS is returned to the sender, 
> the receiver is filled with the data.

If you're using a callback, when the callback is called with
GM_SUCCESS, yes. Your langauge was somewhat ambiguous...

> In particular if a bunch of nodes do directed sends, wait for GM_SUCCESS 
> as above (and below) and then everyone makes a barrier call then, after 
> the barrier returns, everyone can be sure that all the data to be received 
> is in place. Note that this allows compuation to be done while CPU is busy 
> with something else (assuming send tokens, etc. are managed, etc.).

You can also use extra sends for flags, which is cheaper than a
barrier for pairwise synchronization. Remember that GM delivers in
order. Cray SHMEM users tended to mostly use the global barrier method
(hm, that almost sounds obscene), but the fastest method was to use
pairwise synchronizations.

greg