On Aug 11, 2005, at 12:13 PM, Michael Lees wrote:
> Reading further into the article there is something about the IBM mpi
> implementation
>
> " The default interval for handling communication calls is 400
> milliseconds."
>
> Does lam have a similar default?
No. In the lamd RPI, the lamd is a separate process and it simply does
blocking reads on sockets. So it wakes up a) as soon as there is data
to read, and b) when the OS schedules it to run.
> And do I understand this correctly?
> Is this the time between making subsequent checks on a blocking
> communication?
> ie., in MPI_Recv does it check for matching send and then sleep for 400
> miliseconds?
I can't speak for how IBM MPI works, but my *GUESS* is that the 400 ms
interval is for asynchronous progress -- not synchronous progress.
I.e., if you're blocking in MPI_Recv, there's likely no reason to block
for an additional 400 ms; it should check immediately if there's
anything to receive.
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
