ON JOB ASSIGNMENT FOR A PARALLEL SYSTEM OF PROCESSOR SHARING QUEUES

Interest in the job assignment problem for parallel queues has been recently stimulated by research in the area of load balancing in distributed systems, where one is concerned with assigning tasks or processes to processors in order to achieve optimal system performance. However, most of the studies found in the literature refer to a system of parallel queues with FCFS service discipline, while it is well known that the processor sharing (PS) service discipline is often a better model for CPU scheduling in time-shared computer systems. In this paper, we underline some interesting peculiarities of the assignment problem with PS queues as compared to the usual case of the FCFS systems. Also, we propose an approach to the design of assignment algorithms which, in this case, produces solutions performing better than the well-known join-the-shortest-queue (3SQ) assignment rule. However, the algorithms obtained following this approach require more information about the nature of the state of the queueing system. We then illustrate the robustness of the JSQ policy with respect to the nature of the offered load and with respect to system nonhomogeneities, when we consider parallel systems of PS queues. Such properties, together with the cost of the overheads involved in the implementation of more complex policies, seems to indicate that the JSQ policy, although not necessarily optimal, offers a very good solution to the job assignment problem for PS parallel systems. © 1990 IEEE