Scheduling Heterogeneous Systems Using Relative Latencies

Scheduling of parallel modules of an application may produce a significant impact on the performance. The problem of finding optimal schedule is, however, NP-complete. The diversity of the processing elements adds to the complexity that is addressed by presenting heuristic algorithms. This paper presents a novel heuristic Relative Latency-based Scheduling (RLS) for producing schedules for heterogeneous systems. The suggested approach makes use of processing capabilities of the existing processors for diverse number of successors in a task graph while making scheduling decisions. The experiments have been performed on a large number of graphs using different topologies including fork-join, Laplace equation solver, Cholesky decomposition and random graphs. The RLS variants produce better results with a significant difference in schedule length as compared to the well-known HEFT and LMT strategies used for scheduling heterogeneous systems.