Performance analysis of repairable cluster of workstations

Summary form only given. In this paper, an analytical model is developed and used to evaluate average response time of repairable cluster of workstations. Employing the queuing theory, closed form solutions for the response time of fault tolerant clusters of workstations are obtained. Workstations (nodes) in the cluster are divided into two sets: active set and backup set. Fault tolerance is achieved by having a set of active nodes replicate their services at a set of backup nodes. Active nodes, periodically, checkpoint their status on the backups. If an active node fails, one of the backups takes over and joins the active set. Two immediate repair mechanisms are considered to repair faulty nodes in the system. In addition to their closed form formats, the analytical results presented in this paper have several advantages over those presented in the previous work. Unlike previous work, there is no need any more to solve a set of recursive equations and the results reveal much of the characteristics of the system.

1 Introduction

Many systems are required to operate fault-free for a given period of time. Such systems are fault tolerant since they can continue to function even when one or more of its components has failed. Redundancy is often used to achieve high reliability. Over the last decade, the landscape of High Performance Computing (HPC) has changed drastically due to number of significant developments in the speed of the processors and the speed of the communication networks. With this development, it has become possible to build an economical distributed system consisting of powerful workstations interconnected through a high-speed communication link to replace an expensive special-purpose supercomputer. Cluster of Workstations (COWs), or network-based multi-computers, and Massively Parallel Systems (MPPs) are the most prominent distributed-memory systems to replace supercomputers.