Feature request: STARPU_MPI_REDUX starpu_data_access_mode (for inter-node reduction only)?
Dear StarPU developers,
As discussed with @furmento this morning, following her suggestion (thanks!), I am posting this issue to discuss a feature request for a STARPU_MPI_REDUX
starpu_data_access_mode. The goal would be to perform a reduction between processes (typically inter-node) but not within workers of multiple nodes (no intra-node reduction).
Currently, with the STARPU_REDUX
mode only, one important issue is the fact that we may have as many copies of a STARPU_REDUX data on a node as there are cores. Note - if this may help - that it was suggested by @thibault (cf. post-scriptum) that a STARPU_MPI_REDUX
flag that would be only interpreted by the MPI layer might be one way to go.
Thanks much in advance once again for your support,
With best regards,
@AntJego, @abuttari, @guermouc and Manu
PS: This post follows an email discussion on April 23, 2020. I am ccing the recipients @thibault @aumage