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, leading to the release of StarPU 1.1. We have extended our lightweight DSM to support out-of-core scheduling over disks. We have finished integrating StarPU with SimGrid and obtained very accurate simulated times, which allows to experiment scheduling heuristics without having to actually execute the application on the target platform, thus tremendously reducing experimentation time and resource consumption.

We have modularized the scheduling part of StarPU , which permits to create complex schedulers by assembling simple scheduling components. This will allow theoreticians to work on writing the simple scheduling components without having to deal with the technical parts of the scheduling, performed in other scheduling components.

We have also collaborated with various research project to leverage the potential of StarPU : for instance, the PaStiX sparse matrix solver was ported over StarPU , so that we improved the dynamic task and management for applications with such fine-grain task size. This resulted with fair-enough performance on CPUs, compared to the hand-optimized static scheduler of PaStiX, and very promising performance on CPUs + GPUs. EADS ported its sparse hmatrix solver over StarPU , and we collaborated to work on adding StarPU support for communicating sparse data over MPI.