[Starpu-devel] Some interrogations about MPI and StarPU

[Xavier Lacoste <xavier.lacoste@inria.fr>]

Hello,

I have some questions to ask about MPI handling in StarPU :

I forgot the to call the starpu_mpi_cache_flush_all_data() at the end of my algorithm, before the wait_fo_all() :

So i suspected this to be responsible of the memory overhead (a test case won't end with starpu because of memory excedeed while it is OK with native \pastix{} scheduler)

I rerun my tests with the flush and got similar results as without, sometimes better, sometimes worst but always close. And the failing run still fails.

So my first question is : 

- is the starpu_wait_for_all() also posting the  starpu_mpi_cache_flush_all_data() ? That could explain the similar performances

- 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 ?

- When I insert a task invoving a data reception, will starpu alloacte the receiving buffer and post MPI_Irecv or will he probe until the data is here before allocating and receiving the data ?

If it is the first case it could explain the memory overhead as I insert all my receiving tasks at the beginning of the algorithm.

Regards,

XL.

----------------------------------------

Xavier Lacoste

xavier.lacoste@inria.fr

INRIA Bordeaux Sud-Ouest

200, avenue de la Vieille Tour

33405 Talence Cedex

Tél : +33 (0)5 24 57 40 69

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail