Parallel job scheduling on heterogeneous networks of multiprocessor workstations

grantor: University of TorontoParallel job scheduling has been heavily studied on various architectures, from vector supercomputers to networks of workstations. A common assumption of this previous work has been that all processors in the system have equal processing power. We investigate parallel job scheduling on 'heterogeneous ' networks of multiprocessor nodes, where the machines in the network can vary in 'processor speed' and 'node size' (number of processors per node). Two dynamic scheduling policies and two adaptive scheduling policies are simulated in this framework. We provide some evidence that a network's ' coefficient of variation' (standard deviation over mean) of processor speed or node size is a good predictor of job mean response time for these algorithms. As well, processor-speed heterogeneity is found to have a greater effect on job mean response time than node-size heterogeneity. In some cases, the adaptive algorithms can outperform the more flexible dynamic algorithms due to job imbalances created by dynamically reassigning processors.M.Sc