Global backfilling scheduling in multiclusters

Abstract

There exist a centralized scheduler structure and a distributed scheduler structure at multiclusters scheduler. The current centralized scheduler structure requires exact knowledge of scheduling activities about geographically distributed clusters. However, such timely and frequent information dissertation can not be supported by existed network infrastructure. In order to make the centralized scheduler structure realistic, this paper investigates the global scheduler at muliclusters without timely and frequently information distribution. This global scheduler is based on the backfilling scheduling policy and the global scheduler tries to find holes at clusters for queued jobs at other cluster. This global backfilling scheduling policy is evaluated by using real workload trace driven simulation. The simulation results show that proposed policy consistently outperforms the independent site execution. © Springer-Verlag 2004.