Data-bandwidth-aware Job Scheduling in Grid and Cluster Environments

This paper introduces techniques in scheduling jobs on a master/workers platform where the bandwidth is shared by all workers. The goal is to minimize the total makespan. The jobs are independent and each job requires a fixed amount of bandwidth to download input data before execution. The master can communicate with multiple workers simultaneously, provided that the bandwidth used by the master and the workers do not exceed their bandwidth limits. We proposed two models for this limited-bandwidth problem. If the data transfer cannot be interrupted, then we prove that the scheduling problem is NP-complete. Nevertheless we propose heuristic algorithms and experimentally test their performance. If the data transfer can be interrupted, we propose an algorithm that produces optimal makespan. The algorithm is based on a binary search on the completion time, and an efficient feasibility verification process for a given completion time.