Section: New Results

Task scheduling with memory constraints

When dealing with larger and larger datasets processed by task-based applications, the amount of system memory may become too small to fit the working set, depending on the task scheduling order. The ROMA Inria team proposed a heuristic to introduce additional dependencies to the task graph enough so that any scheduling order will meet the memory constraint, while avoiding to extend the critical path length. On the other hand, banker algorithms allow to achieve this online, within the task scheduler, but do not have an overall view of the task graph, and may thus severely increase the critical path. We have thus started to design a collaboration between visionary heuristics which take a global but coarse view of the task graph, and online heuristics which have a local but precise view of the task graph.