Queueing network models for load balancing in distributed systems

Abstract

In distributed systems, load balancing can improve efficiency by migrating jobs from heavily loaded to lightly loaded sites. In this paper we present a method for optimal load allocation in a statistic environment. A queueing network model is used to evaluate response time; and mathematical programing techniques are used to find the load allocation that minimizes average response time. The method is not proposed as a substitute for dynamic, heuristic load balance policies; rather, it is preceived as a useful tool for resource allocation and capacity planning in distributed systems, and as a promising complement to dynamic policies in hibrid load balance strategies. The method can handle very general classes of problems, including: distinct classes of jobs, multitasking within each job, and; jobs with spawned tasks. Several examples illustrating these applications are reported.