there is no buffering in MPI messages (ok, some internal buffering may happen, but it is intentionally reduced to a minimum).
An MPI_Send blocks until the output buffer can be reused. That means that for large messages, it will block at least until the matching receive is posted, and approximately until the transfer is complete.
You may want to look at Bsend and MPI_Buffer_attach, but do not go there, except if you really feel a need for it, this is detrimental to performance.
Le 2 mai 2013 à 15:28, John Chludzinski <john.chludzinski_at_[hidden]> a écrit :
> I assume there are setting for this?
> On Thu, May 2, 2013 at 3:23 PM, John Chludzinski <john.chludzinski_at_[hidden]> wrote:
> If I'm using MPI_Send(...) and MPI_Recv(...) in a producer/consumer
> model and choose not to buffer messages internally (in the app),
> allowing them to acumulate in the MPI layer, how large of an MPI
> message queue can I expect before something breaks?
> users mailing list
* Dr. Aurélien Bouteiller
* Researcher at Innovative Computing Laboratory
* University of Tennessee
* 1122 Volunteer Boulevard, suite 309b
* Knoxville, TN 37996
* 865 974 9375