AN IMPLEMENTATION OF CONCURRENT GANG SCHEDULER FOR PC CLUSTER SYSTEMS

摘要

In this paper, we describe an implementation of Concurrent Gang scheduler for PC cluster systems. Recently, PC cluster systems have become popular because of their cost-effectiveness and flexibility. There is a demand, however, for these systems to be still cheaper. We have designed and implemented a PC cluster system, called FireCluster, which adopts the IEEE 1394 for interconnecting nodes as a solution for achieving a low-cost cluster system. Parallel job scheduling is important to use modern multiprocessor parallel computers efficiently. Although gang scheduler is suitable for cluster systems that provide a user-level communication library, big problem is that the overhead of communication for synchronization by a scheduler may be high in low-cost cluster systems in comparison with systems that use high-cost networks such as Myrinet or SCI. We focus on scheduling based on Concurrent Gang and implement a Concurrent Gang scheduler for PC cluster systems. Experimental results show that our scheduler can synchronize with very low communication overhead.