Evaluating Dynamic Task Scheduling in a Task-Based Runtime System for Heterogeneous Architectures

Heterogeneous parallel architectures present many challenges to application developers. One of the most important ones is the decision where to execute a specific task. As today's systems are often dynamic in nature, this cannot be solved at design time. A solution is offered by runtime systems that employ dynamic scheduling algorithms. Still, the question which algorithm to use remains. In this paper, we evaluate several dynamic scheduling algorithms on a real system using different benchmarks. To be able to use the algorithms on a real system, we integrate them into a task-based runtime system. The evaluation covers different heuristic classes: In immediate mode, tasks are scheduled in the order they arrive in the system, whereas in batch mode, all ready-to-execute tasks are considered during the scheduling decision. The results show that the Minimum Completion Time and the Min-Min heuristics achieve the overall best makespans. However, if additionally scheduling fairness has to be considered as optimization goal, the Sufferage algorithm seems to be the algorithm of choice.