Threshold policies for load sharing in heterogeneous systems

The authors adopt a queuing-theoretic approach to analyzing different threshold policies in the context of a particular model: the single-queue multiple-server (SQMS) model. Computing nodes are represented as servers and the authors consider the general case where they may have different service rates. This is a model of a centrally scheduled system as the entire system work-load goes through a common dispatcher. Jobs cannot be preempted once assigned to servers, and the communication cost is assumed to be included in the processing cost. A class of policies called threshold policies is studied, wherein scheduling decisions are based on whether or not a defined threshold is exceeded. An efficient heuristic for determining the thresholds is presented and its performance is compared with other load sharing policies. The heuristic is found to perform quite well compared with other policies.<
>