Section: New Results

Task scheduling over heterogeneous architectures

We continued our work on extending StarPU to master exploitation of Heterogeneous Platforms through dynamic task scheduling, with a now-imminent release of StarPU 1.2.

We have improved the simulation support with SimGrid , to augment the accuracy of the simulated execution according to the hardware capabilities [30] .

We have collaborated with various research projects to leverage the potential of StarPU . We have improved the support for the PaStiX and QR-Mumps sparse matrix solvers, thus obtaining competitive performance on CPUs and on CPUs+GPUs [25] . We have improved the MPI communication engine of StarPU to get better performance with the EADS hmatrix solver.

We have obtained very good performance and scalability with a Cholesky factorization distributed over a cluster of 144 heterogeneous nodes hosted at CEA.

We have studied the theoretical performance bound that can be achieved for the Cholesky factorization, reproduced the performance of a theorically optimal scheduled, shown that the classical HEFT heuristic is far from it, that more application-specific heuristics allow to get performance closer to the peak, and that the peak is not reachable with simple heuristics, because it requires non-trivial task order inversions.

In relationship with the ADT K'Star effort of building the Klang-OMP OpenMP compiler and putting together the Kastors benchmark suite, StarPU has been extended to provide an OpenMP-enabled runtime support for Klang-OMP . In particular, the StarPU OpenMP Runtime Support implements preemptible tasks required for OpenMP, using the concept of continuations, while maintaining interoperability with StarPU regular, non-blocking tasks, and while preserving the heterogeneous, performance model-based scheduling capabilities of StarPU.

The Klang-OMP C/C++ OpenMP compiler co-developed with Inria Team MOAIS enables plain OpenMP applications to run un-modified on top of the StarPU runtime system, thus significantly increasing the performance portability potential of StarPU.