Improving task scheduling with parallelism awareness in heterogeneous computational environments

Abstract

Task scheduling is a key function for executing tasks in heterogeneous computational environments, efficiently. While the available computing resources are not fully used when applying existing scheduling methods as they consider that a task is executed on one single core or on a server without parallel tasks by assuming that the task exhausts the server. Therefore, in this paper, we focus on the problem of executing tasks with deadline constraints with parallelism awareness where the parallel degree of each task can be tuned between one and its maximum according to the available cores of the server it assigned to during its execution. We first model the problem as an optimization problem maximizing the overall utilization of servers, and propose a set of scheduling methods with parallelism awareness (SPA), each of which iteratively allocates as much resources and as soon as possible to the assigned task with the earliest deadline on a server, based on existing scheduling algorithms, and present two SPA instances to illustrate the implement of SPA. Experiment results show a great performance improvement in various aspects, e.g., resource utilization, task violations, finish time, and energy efficiency, when executing tasks heterogeneous computational systems using SPA.