Estimating a workstation CPU assignment with the DYPAP monitor

Abstract

The task of finding the least loaded node on a cluster or Grid computing system is necessary in load balancing and scheduling algorithms. Therefore, the lint step in these algorithms is usually a prediction of the CPU assignment or availability for a new task on each system node (inversely related to its workload). This paper presents the DYPAP monitor; a real-time monitoring utility responsible for periodically characterizing a node state and computing with this information the dynamic process assignment prediction (DYPAP) model. The goal of this model is to predict the CPU assignment of a workstation from measurements, without a priori knowledge about its state. An extensive set of experiments is presented to validate the DYPAP model and to evaluate the monitor performance in all kind of situations.