On Performance Evaluation of a Slackness Option for the Self-Tuning dynP Scheduler

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 an enhancement to the decision process of the self-tuning dynP scheduler. Adding slackness means, that the currently used policy is virtually improved by a given percentage. This prevents rapid and consecutive policy switches, which might be induced by users for cheating the scheduler. We use discrete event simulations to evaluate the performance. As job input for driving the simulations we use original traces from real supercomputer and cluster installations. To increase the workload to be processed by the scheduler, the average interarrival time is decreased with the shrinking factor. The evaluation of the slackness enhancement shows, that if slackness is applied to the simple and advanced deciders both are equal in their decisions and generate the same performance. This is due to the fact, that in the decision process cases with two equal performing basic policies do no longer occur. The results show, that it depends on the trace, if, and how much slackness is beneficial. In general, the performance of the self-tuning dynP scheduler is increased by applying small slackness values. The performance benefit of slackness is most evident for the CTC trace.