Using On-line Job Monitoring for Dynamic Load Balancing of Parallel Applications

While the significance of the communication costs for the performance of parallel applications is well known, these costs are hardly considered in the known resource management systems for workstation clusters. In this paper we present the Cluster Search Algorithm (CSA) for dynamic remapping of parallel applications in clusters. The novelty of our approach is that the communication behavior of the application does not have to be known a priori. The decisions are made dynamically at run time according to the actually observed communication behavior of the application and available resources. We present first simulation results and an overview over an actual implementation.