Enhancements to the Decision Process of the Self-Tuning dynP Scheduler

摘要

The self-tuning dynP scheduler for modern cluster resource management systems switches between different basic scheduling policies dynamically during run time. This allows to react on changing characteristics of the waiting jobs. In this paper we present enhancements to the decision process of the self-tuning dynP scheduler and evaluate their impact on the performance: (ⅰ) While doing a self-tuning step a performance metric is needed for ranking the schedules generated by the different basic scheduling policies. This allows different objectives for the self-tuning process, e. g. more user centric by improving the response time, or more owner centric by improving the makespan. (ⅱ) Furthermore, a self-tuning process can be called at different times of the scheduling process: only at times when the characteristics of waiting jobs change (half self-tuning), i. e. new jobs are submitted; or always when the schedule changes (full self-tuning), i.e. when jobs are submitted or running jobs terminate. We use discrete event simulations to evaluate the achieved performance. As job input for driving the simulations we use original traces from real supercomputer installations. The evaluation of the two enhancements to the decision process of the self-tuning dynP scheduler shows that a good performance is achieved, if the self-tuning metric is the same as the metric used measuring the overall performance at the end of the simulation. Additionally, calling the self-tuning process only when new jobs are submitted, is sufficient in most scenarios and the performance difference to full self-tuning is small.