Demand-based scheduling priorities for performance optimisation of stream programs on parallel platforms

This paper introduces a heuristic-based scheduler to optimise the throughput and latency of stream programs with dynamic network structure. The novelty is the utilisation of positive and negative demands of the streamcommunications. It is a centralised approach to provide load balancing for stream programs with dynamic network structures. The approach is designed for shared-memory multi-core platforms. The experiments show that our scheduler performs significantly better than the reference implementation without demand considerations.