An Efficient Execution Model for Reactive Stream Programs

Stream programming is a paradigm where a program is structured by a set of computational nodes connected by streams. Focusing on data moving between computational nodes via streams, this programming model fits well for applications that process long sequences of data. We call such applications reactive stream programs (RSPs) to distinguish them from stream programs with rather small and finite input data. In stream programming, concurrency is expressed implicitly via communication streams. This helps to reduce the complexity of parallel programming. For this reason, stream programming has gained popularity as a programming model for parallel platforms. However, it is also challenging to analyse and improve the performance without an understanding of the program’s internal behaviour. This thesis targets an efficient execution model for deploying RSPs on parallel platforms. This execution model includes a monitoring framework to understand the internal behaviour of RSPs, scheduling strategies for RSPs on uniform shared-memory platforms; and mapping techniques for deploying RSPs on heterogeneous distributed platforms. The foundation of the execution model is based on a study of the performance of RSPs in terms of throughput and latency. This study includes quantitative formulae for throughput and latency; and the identification of factors that influence these performance metrics. Based on the study of RSP performance, this thesis exploits characteristics of RSPs to derive effective scheduling strategies on uniform shared-memory platforms. Aiming to optimise both throughput and latency, these scheduling strategies are implemented in two heuristic-based schedulers. Both of them are designed to be centralised to provide load balancing for RSPs with dynamic behaviour as well as dynamic structures. The first one uses the notion of positive and negative data demands on each stream to determine the scheduling priorities. This scheduler is independent from the runtime system. The second one requires the runtime system to provide the position information for each computational node in the RSP; and uses that to decide the scheduling priorities. Our experiments show that both schedulers provides similar performance while being significantly better than a reference implementation without dynamic load balancing. Also based on the study of RSP performance, we present in this thesis two new heuristic partitioning algorithms which are used to map RSPs onto heterogeneous distributed platforms. These are Kernighan-Lin Adaptation (KLA) and Congestion Avoidance (CA), where the main objective is to optimise the throughput. This is a multi-parameter optimisation problem where existing graph partitioning algorithms are not applicable. Compared to the generic meta-heuristic Simulated Annealing algorithm, both proposed algorithms achieve equally good or better results. KLA is faster for small benchmarks while slower for large ones. In contrast, CA is always orders of magnitudes faster even for very large benchmarks.