On the Performance of a Clustering-based Task Scheduling in a Heterogeneous System

Abstract

Recent task scheduling algorithms for a generalized workflow job in heterogeneous system adopt list-based scheduling. In those algorithms, the response time cannot be effectively reduced if the given workflow job is data-intensive. If the workflow job is computationally intensive, an attempt is made to assign tasks to many processors, which can lead to resource starvation. To this end, a task scheduling algorithm that is based on clustering tasks, called CMWSL (Clustering for Minimizing the Worst Schedule Length) has been proposed. In CMWSL, the lower bound of the assignment unit size for each processor is derived in order to suppress the total number of executing processors for effective use of processors. After the lower bound is derived, the processor as a assignment target is determined and then the assignment unit as a task cluster is generated. As a final phase of CMWSL, task ordering is performed for every assigned task. In this paper, we compare several task ordering methods in CMWSL in a real environment to find the best task ordering policy.