Buffer minimization in pipelined SDF scheduling on multi-core platforms

With the increasing number of cores available on modern processors, it is imperative to solve the problem of mapping and scheduling a synchronous data flow graph onto a multi-core platform. Such a solution should not only meet the performance constraint, but also minimize resource usage. In this paper, we consider the pipeline scheduling problem for acyclic synchronous dataflow graph on a given number of cores to minimize the total buffer size while meeting the throughput constraint. We propose a two-level heuristic algorithm for this problem. The inner level finds the optimal buffer size for a given topological order of the input task graph; the outer level explores the space of topological order by applying perturbation to the topological order to improve buffer size. We compared our proposed algorithm to an enumeration algorithm which is able to generate optimal solution for small graphs, and a greedy algorithm which is able to run on large graphs. The experimental results show that our two-level heuristic algorithm achieves near-optimal solution compared to the enumeration algorithm, with only 0.8% increase in buffer size on average but with much shorter runtime, and achieves 38.8% less buffer usage on average, compared to the greedy algorithm.