SCAPE: HW-Aware Clustering of Dataflow Actors for Tunable Scheduling Complexity

Abstract

This paper introduces a fast method to generate high performance parallelized code from a dataflow specification of an application. Dataflow Models of Computation (MoCs) are efficient programming paradigms for expressing the parallelism of an application. Traditionally, mapping and scheduling methods for dataflow MoCs rely on complex graph transformations to explicit their parallelism which can result in complex graph for embarrassingly parallel applications. For such applications, state-of-the-art mapping and scheduling techniques are prohibitively complex, while the exposed parallelism often exceeds the parallel processing capabilities of the target architecture. We propose SCAPE, an automated method to control the complexity of the pre-scheduling graph transformation by using information from the architecture and application models. By decreasing the complexity of the graph, the mapping scheduling task is accelerated at the potential expense of the produced schedule. Our method offers a limited and controlled decrease of the schedule quality while enabling mapping and scheduling execution time between 1 and 2 orders of magnitude faster than state-of-the-art techniques.