Section: New Results

Progress threads placement for MPI Non-Blocking Collectives

MPI Non-Blocking Collectives (NBC) allow for communication overlap with computation. A good overlapping ratio is obtained when computation and communication are running in parallel. To achieve this, each MPI task generates a progress thread to manage communication tasks. The progression of these communications requires regular access to the processors. These threads compete with each other and with MPI tasks. In order to run threads with minimal disruption, we bound the progress threads on free cores when it is possible. Then, we showed that folding all progress threads on very few cores does not work for tree algorithms. The number of communication generated are too important. The solution that we propose is to perform a number of levels (S) of the dependency tree on MPI tasks. We get a reasonable execution time (less than compute time + communication time) while reserving fewer cores for progress threads. All these methods have been implemented in the MPC framework, which contributes to its development.