Replica-Aware Job Scheduling in Distributed Systems

This paper proposes an effective replica-aware scheduling algorithm for independent jobs in Grid and distributed systems. The proposed algorithm considers not only the execution time of jobs but also the location and transfer time of data and data replica that these jobs require. We propose a cost model to estimate the starting time and earliest completion time of a job and its associated data (original or replicated). Based on the estimated time, the scheduling algorithm finds a proper execution sequence for the jobs and the data with the goal to minimize the makespan of the jobs. Our experiment results demonstrate that the proposed algorithm is scalable and outperforms a random job selection strategy. We also show that the proposed algorithm performs well compared to a conservative theoretical lower bound, with performance within 15% of the lower bound on average and within 40% in the worst case.