Section: Overall Objectives

Highlights of the Year

The PaStiX solver is now able to handle efficiently multiple GPU accelerators using runtime systems (StarPU or PaRSEC ). On the Plafrim machine, one GPU card can provide almost the same performance than 12 cores and we get a good scalability while mixing multicores and upto 3 GPUs accelerators.

The first implementation of the Fast Multipole method over a runtime system has been developed in the context of the FastLA associated team. The main outcome of this work will be published in a paper to appear in the SIAM SISC journal.