Section: New Results

High Performance methods for solving wave equations

Last year we optimized the DG code implemented in the DIVA platform of Total by reducing the number of communications between each processors. This optimization, coupled with the use of Hybrid MPI and OpenMP parallel programming has allowed to prove the scalability of the code up to 512 cores. We are now planning to extend these tests up to 4000 cores. However, preliminary results emphasized the limitations due to low level issues such as threads placement, data communications and cache utilization. Therefore, we are now considering the implementation in DIVA of task programming libraries such as StarPU (http://runtime.bordeaux.inria.fr/StarPU/ or DAGuE (http://icl.cs.utk.edu/dague/index.html ). These libraries handle the low level issues directly at the runtime level and allow the programmer to focus on the algorithm itself. They are also provide a valuable help to improve the portability of the code from one architecture to another, which will allow us to port DIVA on heterogeneous architectures such as CPU/GPU and Intel Xeon Phi. We have already coupled DIVA and DAGuE on the Symmetric Multiprocessor System (SMP) of Plafrim (https://plafrim.bordeaux.inria.fr ) and compared the performance of the code with MPI and an OpenMP implementations [66] , [64] .