A comparison of scheduling algorithms for multiprocessortasks with precedence constraints

Many parallel applications from scientic computing show a modular structure and are therefore suitable for the multiprocessortask programming model with precedence constraints. This programming model has been shown to yield better results than a pure dataparallel or a pure taskparallel execution on distributed memory platforms in many cases. The efficient execution of multiprocessortask programs requires an appropriate schedule, which takes the structure of the application and the performance characteristics of the target platform into account. Many heuristics and approximation algorithms have been proposed to fulfil this scheduling task. In this paper we consider popular scheduling algorithms that have been implemented in a scheduling toolkit. Specifically, we introduce Allocation-and-Scheduling-based algorithms and compare their runtime for large task graphs consisting of up to 1000 nodes and target systems with up to 256 processors. Furthermore we consider the quality of the produced schedules and derive a guideline describing which scheduling algorithm is most suitable in which situation.