Engineering grid applications and middleware for high performance

Abstract

Meeting QoS objectives of applications while maintaining high system utilization is a challenging task in multi-institutional Grids. In this paper, we effectively engineer Grid applications and resource management middleware for achieving user satisfaction and high resource utilizations. The paper presents a complete framework based on advance reservations (ARs) for resource management in Grids. The framework is capable of providing QoS guarantees to applications while maintaining high resource utilizations. The paper focuses on the scheduling component of the framework and presents a novel heuristic-based algorithm, Grid Scheduling with Deadlines (GSD), for an NP-Complete problem of scheduling ARs with laxities on a shared resource. GSD can be configured with the help of pluggable strategies to adapt to various workload conditions and needs of the system. The paper studies with the aid of an extensive set of experiments the effect of various workload and system parameters on system performance. It is not always possible to accurately predict the runtimes of the jobs. The paper discusses the impact of error in user-estimated runtimes on system performance and investigates strategies to avoid substandard performance resulting from such inaccuracies. Experimental results demonstrate the efficacy of our methodology.