Coupling the DG code with task programming libraries

Participants : Lionel Boillot, Emmanuel Agullo, George Bosilca, Henri Calandra.

The parallelization of the original code is based on a preliminary step of domain decomposition and then on the use of the MPI (Message Passing Interface) library. It is a common choice which works pretty well in most of the classical architectures. However, the parallel efficiency is not optimal and the performance decreases in hybrid architectures. Indeed, we know the number of operations that each sub-domain has to performed but this does not give us the exact time that the computations require. The cluster heterogeneity leads to various automatic optimizations (memory cache, parallel capability, …) which are difficult to measure. We have decided to tackle this problem by modifying the parallelism with the use of task programming. We have thus rewritten the DIVA algorithm in a graph of tasks without using the MPI library and we have left to the runtime PaRSEC the choice of when and where to execute each task. The numerical experiments we have performed have confirmed the significant improvement of the parallel efficiency on different architectures like ccNUMA machines or Intel Xeon Phi co-processors. Moreover, the proposed solution is portable on these architectures, this means that none or few modifications are required in the code, allowing to focus on algorithmic aspects in order to preserve the performance. These results have been presented to the EAGE HPC workshop and to the HPCC IEEE conference within a paper have been accepted.