A survey of load sharing in networks of workstations

This paper is a survey of existing policies and mechanisms for load sharing in loosely-coupled distributed computing systems, where user machines are personal workstations interconnected by a local area network. We are interested only in centralized operating systems providing mechanisms for remote process communication, thus we do not study distributed operating systems in which load balancing and process migration may be provided by a network-wide virtual memory mechanism. We define load sharing, load balancing, non-preemptive migration and pre-emptive migration, and we discuss the goals of load sharing and load balancing strategies related to process scheduling. We argue against the usefulness of load balancing strategies in the context of networks of workstations. A load sharing algorithm is composed of three parts, namely, a location policy, an information policy and a transfer policy. We review the different location policies, information policies and transfer policies that have been proposed in the literature. We discuss their ability to take personal use of workstations into account, and we compare them with respect to the performance that can be really obtained in the context of networks of workstations. We show that only some policies are efficient in such a context. Thereafter, we present the mechanisms proposed for supporting the policies, and discuss them with respect to network interfaces, file system design, machine heterogeneity and program interactivity. The question of pre-emptive versus non-pre-emptive migration is addressed, and we argue that pre-emptive migration does not provide substantial benefits in a network of workstations. Most of the implemented load sharing systems described In the literature are presented, and finally some perspectives in this research area are described.