Abstract
Coscheduling has been gained a resurgence of interest as an effective technique to enhance the performance of parallel applications in multi-programmed clusters. However, existing coscheduling schemes do not adequately handle priority boost conflicts, leading to significantly degraded performance. To address this problem, in our previous study, we devised a novel algorithm that reorders the scheduling sequence of conflicting processes based on the rescheduling latency of their correspondents in remote nodes. In this paper, we exhaustively explore the design issues and implementation details of our contention-aware coscheduling scheme over Myrinet-based cluster system. We also practically analyze the impact of various system parameters and job characteristics on the performance of all considered schemes on a heterogeneous Linux cluster using a generic coscheduling framework. The results show that our approach outperforms existing schemes (by up to 36.6% in avg. job response time), reducing both boost conflict ratio and overall message delay.
