Orchestrating of Stream Programs by Genetic Algorithm

Abstract

Streamprogrammingonparallelhardwarehasbecomeubiquitous. Inthismodel,aprogram is executed on stream graph which consists of a set of actors performing different functions communicating through edges. It is difficult to schedule actors onto multicores balancing the workload because the assigned actor load can be overshadowed by the communication overhead of the edges. The proposed algorithm presents a heuristic technique to schedule the actors on multicores that balances the workload. Although the problem is time consuming and has much more computational overhead, this paper presents an efficient and effective technique that provides nearly optimal solution with minimum cost. We present a random partitioning steps based on Genetic Algorithm that assigns a set of actors to multicore processors in such a way that minimize the maximum total processing time on any processor. For facilitating understanding, an example of a stream graph is presented to visualize the chronological development of genetic algorithm. Our algorithm for scheduling actors in multicore is highly efficient and its results are accurate for standard number of actors and close to the optimal solution for larger number. For a range of stream graph considering the communication overhead a speedup of 3.87x on 4-Core, 5.72x on 6-Core and 7.59x on 8-Core architecture is achieved compared to a single core.