Efficient Task Assignment on Heterogeneous Multicore Systems Considering Communication Overhead

Abstract

AbstractThis paper addresses task assignment problem on heterogeneous multicore systems with time constraint considering communication overhead. Processing cores in a heterogeneous system considered in this paper are grouped into clusters according to core types. Therefore, clusters have different computation capabilities. Communication links among various clusters have different communication capacities as well. The goal of heterogeneous task assignment problem is to minimize the total system cost for allocating a set of given tasks with data dependencies to a group of heterogeneous clusters while the time constraint is satisfied. The system cost considered in this paper is related to both execution load and communication load on various clusters and communication links. The general heterogeneous assignment problem is NP-complete. In this paper, we present the ILP formulation for solving the heterogeneous assignment problem. We also propose a heuristic, the Ratio Greedy Assign algorithm (RGA), to solve the problem efficiently for directed acyclic graphs (DAG). According to our experimental results, the Ratio Greedy Assign algorithm generates near-optimal results efficiently for all the benchmarks, while the ILP method cannot find a solution with acceptable computation time for large-sized benchmarks, such as 10-4lattice filter. Compared with a method that assigns all the tasks to a cluster of homogeneous cores, the RGA algorithm reduces the total system cost by 35.1% on average with four heterogeneous clusters. It reduces the cost by 24.6% on average with three heterogeneous cluster.KeywordsTask assignmentMulticore systemILPGreedy algorithm