Section: New Results

Thread placement and threads policy on a multicore machine with NUMA effects.

Threads placement on multicore machine with NUMA effects is inevitable to have better performances. Threads must bind on cores to avoid thread migration and to have better cache locality. MPI non-blocking collectives can generate progress threads to complete communications. These additional threads can disturb computational threads. That is why we have implemented several thread placement algorithms into the MPC framework  [22] . These algorithms allow to dedicate resources only for progress threads. Thus computational threads are not disturb. We test them with our own benchmarks which test all the MPI non-blocking collectives to compare the performances with different thread placement. We observe an improvement when resources are dedicated to progress threads and take NUMA effects into account.

We want to include a mechanism into MPC to specify thread kinds (MPI, OpenMP,...). These mechanism will allow the MPC scheduler to take threads specificity into account to improve the scheduling policy. Our goal is to increase runtime performances considering each type specific needs. We have begun to implement this mechanism.

Several MPC framework bugs have been corrected, thus we contribute to its stability.