Re: [Starpu-devel] Some interrogations about MPI and StarPU

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Xavier Lacoste, le Wed 29 Oct 2014 14:18:51 +0100, a écrit :
> - is the starpu_wait_for_all() also posting the
>  starpu_mpi_cache_flush_all_data() ? That could explain the similar
> performances

No.

> - Would it be better to insert the flush as soon as I know it can be 
> flushed or
> is the final flush_all_data() equivalent ?

It is equivalent when you do not reuse the data after explicitly
flushing it: the flush submitted by flush_all_data() will happen as soon
as all tasks using the data are finished.

> - When I insert a task invoving a data reception, will starpu alloacte the
> receiving buffer and post MPI_Irecv

It will currently allocate the receiving buffer as soon as the task is
inserted. The idea is to make the MPI communication happen before the
task dependencies are available. Benoît suggested waiting for the task
dependencies, but that'll reduce the available parallelism. What Marc is
doing ATM is not submitting all tasks immediately, by waiting explicitly
in the most external loop of its submission algorithm.

With starpu 1.2 where starpu performs its own sync between sender and
receiver, whenever somebody gets the time to implement it, we should be
able to make starpu allocate only once the sender has the data to send.

Samuel