Load sharing in limited access distributed systems

In this paper we examine dynamic load sharing in limited access distributed systems. In this class of distributed systems all servers are not accessible to all sources, and there exist many different accessibility topologies. We focus our attention on the ring topology and provide an analytic model to derive the approximate mean waiting time (our metric of performance). We then consider other limited access topologies and find that rather different interconnection patterns give similar performance measurements. We conjecture that the number of servers accessible to a source is the parameter with the greatest performance impact, in a limited access topology with load sharing. We also introduce another variable called diversity that is indicative of the degree of load sharing and speculate that performance is reasonably insensitive to diversity so long as it is non-zero. Using these conjectures we show how a reasonable estimate of the mean waiting time can be analytically derived in many limited access topologies.