大规模并行计算机作业调度评价

Evaluation of Massive Parallel Computer Job Scheduling

大规模并行计算机的作业调度直接关系到其计算能力的发挥,因而相应的研究具有十分重要的意义。论文通过对国外现有作业调度评价体系的研究,建立了更能反映并行作业特点的作业调度策略评价体系,在此基础上设计并实现了一个作业模拟调度环境。模拟调度环境采用事件驱动的工作模式,支持FCFS、大作业优先、小作业优先、长作业优先、短作业优先、GANG等调度策略。模拟测试结果表明,GANG调度策略优于所有测试的空间共享调度策略;同时在空间共享调度策略中,短作业优先策略和大作业优先策略具有较好的性能。

Since parallel job scheduling in Super-computer was gained recognition as important topic,much research on scheduling strategy and performance evaluation has been done but it is still difficult to choose the most efficient strategy for a given computing environment.In this paper,we improve the existed metrics of parallel job scheduling and implement a simulated environment for job scheduling.The simulating environment supports both space-sharing strategy such as FCFS and time-sharing strategy.In order to present the results of the empirical study of strategy other than the conceptual result,we construct real workload derived from the supercomputer job logs for the environment.Our experimental results demonstrate that GANG scheduler can be more effective than almost all space-sharing strategy especially when the environment has heavy load.The short run-time job first and the big job first strategy can be effective among all space-sharing strategy.