[Starpu-devel] Difference between starpu_mpi_cache_flush() and starpu_mpi_cache_flush_all() ?

[Benoît Lizé <benoit.lize@gmail.com>]

Hello,

I am using StarPU-MPI 1.1, and want to take advantage of the cache flush mechanism that is now available. I was thinking about the following approach:

insert_all_the_tasks();

starpu_mpi_cache_flush_all_data();

This leads me to two questions:

1.

From the code, it seems to me that there is no difference between:

for (int i = 0; i < nhandles; i++) {

  starpu_mpi_cache_flush(MPI_COMM_WORLD, starpu_handles[i]);

}

and

starpu_mpi_cache_flush_all_data(MPI_COMM_WORLD);

provided that all the registered handles are in the "handles" array.

Is it true ?

2.

From what I understand, and assuming that starpu_mpi_insert_task() doesn't block (which AFAIK is true right now), there should not be a difference between flushing the cache as soon as I know that a given handle won't be touched anymore, and using starpu_mpi_cache_flush_all() at the end.

Is it true ?

If so, why does the Cholesky example in mpi/examples/matrix_decomposition/mpi_cholesky_codelets.c do it handle by handle ?

Thank you,

-- 

Benoit Lize