Fuzzy based resource management framework for high throughput computing

We suggest a fuzzy based scheduling policy, which not only considers the distributed ownerships and heterogeneous environments but also handles the dynamic state of the cluster (load etc) more effectively. The algorithm aims at finding the best schedule for the combined set of jobs using fuzzy logic. Two fuzzy membership functions, resource fuzzy set and priority fuzzy set, are computed based on each job's requirements against each available node and user priority and are combined using a convex combination of fuzzy numbers to arrive at a final fuzzy value. This value indicates the extent to which the job can be executed on the given node. The result, of this formulates a score matrix, which is converted to a cost matrix and solved for minimum cost using the Hungarian approach of assignment. One of the important factors in this approach is that of deciding weights for the resources. We argue that the weights for each resource cannot be static. We propose a pairwise comparison of all resources to arrive at the weights for each resource for each job. This framework we feel is highly scalable and adapts to dynamic changes in the cluster. Unlike the other approaches, which primarily address the issue of locating a resource that matches the jobs' requirements, this method addresses the issue of performance/throughput of the cluster also.