PERFORMANCE ANALYSIS OF CONTENTION AWARE MIXED PARALLEL SCHEDULING HEURISTICS ON MULTIPROCESSOR PLATFORMS

Abstract

represented as a Directed Acyclic Graph (DAG). A recent trend has been towards exploiting mixed data and task parallelism which is inherently present in many scientific applications. Most of these algorithms are not suitable for target systems with high contention. This is because they assume idealistic target platform and assume that the communication resources are contention free. Most of these algorithms also assume a fully interconnected network topology. The result is that the produced schedules are not accurate and this results in poor performance of these algorithms. This thesis work aims at modifying the scheduling heuristics to incorporate contention awareness and thus make them suitable for any arbitrary target system topology. The performance of these algorithms, after modification, is analyzed. Experiments were conducted to study the extended algorithms. The obtained results reflected the improvement in schedule lengths by incorporating contention awareness. iii