Analysis of the drop-out rule in probabilistic load sharing in distributed computing systems

It is known that the optimal solution to a probabilistic load-sharing problem in a distributed computing system may require that some of the slower processors in the system are not assigned any processing load. This is known as the drop-out rule. Although this rule has been used quite extensively, a formal exposition of the rule is not available in the literature. The drop-out rule is formally derived, and algorithms to isolate the inactive processors from the active ones, when the service times follow a general distribution and also when they follow an exponential distribution, are presented. The correctness of the algorithm is proved through a rigorous analysis of the drop-out rule. An illustrative numerical example is also provided.