Section: New Results

Leveraging compiler analysis for task scheduling

Polyhedral analysis of task graph submission loops allow to get at compilation time a representation of the task graph, and perform insightful analyses thanks to the obtained overview of the whole task graph. Task scheduling heuristics, on the other hand, usually keep only a limited view over the task graph, to avoid prohibitive algorithmic costs. We have started to collaborate with the CASH Inria team to transfer some of the insights of the compiler to the compiler. We have notably made the compiler automatically compute a cut of the task graph below which the availability parallelism is lower than the capacities of the target hardware. The scheduler can then at that point switch between a heuristic which privileges task acceleration, and a heuristic which privileges the critical path. Only preliminary results have been obtained so far.