Dynamic job scheduling on heterogeneous clusters

Abstract

This paper addresses the problem of scheduling dynamically multi-user and independent jobs on clusters, both homogeneous and heterogeneous. The dynamic behavior means that the scheduler is able to adapt the scheduling when new jobs are submitted and also when processors availability changes. The scheduler has two main features comparing to other solutions: it considers a job as being described by a direct acyclic graph (DAG) and it is able to schedule parallel tasks, when appropriate, instead of the common dynamic mapping approach that assigns an entire job to a processor or a fixed set of processors. The scheduling method is divided in a scheduling strategy and a scheduler algorithm, so that other scheduling algorithms can be incorporated. In this paper two static DAG schedulers for heterogeneous machines are considered. The results show the behavior of the scheduling method for the short completion time of a batch of jobs. These results show better performance when compared to the common schedulers strategies that fix the number of processors per job or assign one processor per job.