Faithful performance prediction of a dynamic task-based runtime system for heterogeneous multi-core architectures

摘要

Multi-core architectures comprising several graphics processing units (GPUs) have become mainstream inrnthe field of high-performance computing. However, obtaining the maximum performance of such heterogeneousrnmachines is challenging as it requires to carefully off-load computations and manage data movementsrnbetween the different processing units. The most promising and successful approaches so far build onrntask-based runtimes that abstract the machine and rely on opportunistic scheduling algorithms. As a consequence,rnthe problem gets shifted to choosing the task granularity, task graph structure, and optimizing thernscheduling strategies. Trying different combinations of these different alternatives is also itself a challenge.rnIndeed, obtaining accurate measurements requires reserving the target system for the whole duration ofrnexperiments. Furthermore, observations are limited to the few available systems at hand and may be difficultrnto generalize. In this article, we show how we crafted a coarse-grain hybrid simulation/emulation of StarPU,rna dynamic runtime for hybrid architectures, over SimGrid, a versatile simulator of distributed systems.rnThis approach allows to obtain performance predictions of classical dense linear algebra kernels accuraternwithin a few percents and in a matter of seconds, which allows both runtime and application designers tornquickly decide which optimization to enable or whether it is worth investing in higher-end graphics processingrnunits or not. Additionally, it allows to conduct robust and extensive scheduling studies in a controlledrnenvironment whose characteristics are very close to real platforms while having reproducible behavior.