Abstract

While it is well known that queues can build up at various points in large scale multiprogramming computer systems, comparatively little is known about the factors which govern the lengths of these queues and their relationship to overall system performance. The first part of this paper is concerned with developing a queueing network model which can be used to study a number of such questions. The model is then applied to a specific problem concerning the proportion of processing requests which should be directed to each of a set of functionally equivalent peripheral processors (e.g., disks and drums) in order to optimize overall system performance. A surprising result is that optimal performance is not attained when queue lengths and processor utilization percentages are equal, but rather when the fastest processor has the longest expected queue and is in effect creating a system bottleneck.