Divisible Load Scheduling on Multi - level Processor Trees

Divisible Load Theory (DLT) is an effective tool for blueprinting data-intensive computational problems. Heuristic algorithms have been proposed in the past to solve for a DLS (Divisible Load Schedule) with result collection on heterogeneous star networks. However scheduling on heterogeneous multi-level trees with result collection is still an open problem. In this thesis, new heuristic algorithms for scheduling divisible loads on heterogeneous multi-level trees (singleand twoinstallment) including result collection are presented. Experiments are performed on both random networks and cluster networks. Results show that scheduling using multi-level trees produces lower solution times compared to the traditional star network in the majority of cases, however efficiency of resources in multi-level trees tends to be lower, i.e., more processors were used. Cluster results with multi-level trees are found to outperform the star when there are enough clusters available to provide good overlap of communication and computation. Experiments on random networks with varying levels of heterogeneity of resources show that multi-level trees outperform star networks in the majority of cases. Experiments were conducted comparing schedules with and without latency costs. The results from all schedules where latency was considered had signifiantly lower solution times and higher efficiency of resources. Overall, scheduling on single-installment multi-level trees in either clusters or random networks had the lowest solution times, but the star had highest efficiency of resources.