Layer-Based Scheduling Algorithms for Multiprocessor-Tasks with Precedence Constraints

A current challenge in the development of parallel applications for distributed memory platforms is the achievement of a good scalability even for a high number of processors. The seatability is impacted by the use of communication operations. e.g. broadcast operations. whose runtime exhibits a logarithmic or linear dependence on the number of utilized processors. The multiprocessor-task programming model can help to reduce the communication overhead, but requires an appropriate schedule for an efficient execution. Many heuristics and approximation algorithms are available for this scheduling task. The choice of a suitable scheduling algorithm is an important factor in the development of multiprocessor-task applications. In this paper. we consider layer-based scheduling algorithms and compare their runtimes for large task graphs consisting of up to 1000 nodes and target systems with up to 256 processors. Furthermore. we introduce an extension methodology to enable scheduling algorithms for independent multiprocessor-tasks to handle precedence constraints.