16:00
18:00

To meet the exascale challenge, parallel programming models tend to abstract the machine details with task-based programming or by exposing the hardware at the software level (network with RDMA capabilities) with the Partitioned Global Address Space (PGAS) model. The Message Passing Interface (MPI) standard proposes a similar communication model to the PGAS model, via the MPI One-Sided Communications. These operations allow a better overlap of communications by computations compared to classic two-sided communications. However, programmers must handle data consistency and synchronisation to prevent data races, which can be a challenging task. Our study proposes two solutions to facilitate the use of one-sided communications in MPI progams. The first is to improve an existing data race detection tool to make it more scalable. The second is a tool that automatically transform two-sided communications into one-sided communications.

English
Amphi LaBRI