An Evaluation of Job Scheduling Strategies for Divisible Loads on Grid Platforms
In this paper, the authors study distributed job scheduling in grid environments when each job is a DL application. The scheduling goal is to minimize the average steady-state job turnaround time. In this context, they identify in which regimes classes of scheduling strategies are efficient, namely for which platforms and which communication to computation ratios. They also quantify what level of global information about the platform is required for efficient scheduling. All their findings are obtained via simulation of wide ranges of application and platform scenarios. Their most significant findings are that the use of grid information is only necessary at high workload, and that at high workload using dynamic information improves performance by around 10% when compared to using static information.