Section: New Results

Parallel Tasks within StarPU

One of the biggest challenge raised by the design of high performance task-based applications on top of heterogeneous accelerator-based machines lies in choosing the adequate granularity of tasks. Indeed, GPUs generally exhibit better performance when executing kernels featuring numerous threads whereas regular CPU cores typically reach their peak performance with fine grain tasks working on a reduced memory footprint. As a consequence, task-based applications running on such heterogeneous platforms have to adopt an intermediate granularity, chosen as a trade-off between coarse-grain and fine-grain tasks. We have tackle this granularity problem via resource aggregation : our approach consists in reducing the performance gap between accelerators and single cores by scheduling parallel tasks over cluster of CPUs. For this purpose, we have extended the concept of scheduling context, which we introduced in a previous work, so that the main runtime system only sees virtual computing resources on which it can schedule parallel tasks (e.g. BLAS kernels). The execution of tasks inside such clusters can virtually rely on any thread-based runtime system, and does not interfere with the outer scheduler. As a proof of concept we allow the interoperability of StarPU and OpenMP to co-manage task parallelism. We showed that our approach is able to outperform the magma, dplasma and chameleon state-of-the-art dense linear algebra libraries when dealing with matrices of small and medium size.