A Graph based approach for Co-scheduling jobs on Multi-core computers

Abstract

In a multicore processor system, running multiple applications on different cores in the same chip could cause resource contention, which leads to performance degradation. Recent studies have shown that job co-scheduling can effectively reduce the contention. However, most existing co-schedulers do not aim to find the optimal co-scheduling solution. It is very useful to know the optimal co-scheduling performance so that the system and scheduler designers can know how much room there is for further performance improvement. Moreover, most co-schedulers only consider serial jobs, and do not take parallel jobs into account. This paper aims to tackle the above issues. In this paper, we first present a new approach to modelling the problem of co-scheduling both parallel and serial jobs. Further, a method is developed to find the optimal co-scheduling solutions. The simulation results show that compare to the method that only considers serial jobs, our developed method to co-schedule parallel jobs can improve the performance by 31% on average.