G-LOMARC-TS: Lookahead group matchmaking for time/space sharing on multi-core parallel machines

Parallel machines with multi-core nodes are becoming increasingly popular. The performances of applications running on these machines are improved gradually due to the resource competition in each node. Researches have found that coscheduling different applications with complementary resource characteristics on the same set of nodes (semi time sharing) may improve the performance. We propose a scheduling algorithm G-LOMARC-TS which incorporates both space and semi time sharing scheduling methods and matches groups of jobs if possible for coscheduling. Since matchmaking may select jobs further down the waiting queue and the jobs in front of the queue may be delayed subsequently, fairness for each individual job will be watched and the delay will be kept within a limited bound. Several heuristics are used to solve the NP-complete problem of forming groups. Our experiment results show both utilization gain and average relative response time improvements of G-LOMARC-TS over other several scheduling policies