Scheduling Real-time Divisible Loads in Cluster Computing Environment

Abstract

The significance of cluster computing in solving massively parallel workloads is tremendous. Divisible Load Theory has proven to be very successful in optimizing the usage of the system resources by partitioning the arbitrarily divisible loads adequately among the cluster nodes. Arbitrarily divisible loads have significant real-world applications in high energy and particle physics. In this thesis, various algorithms for a cluster computing environment are studied including the ones dealing with divisible load theory confirming DLT based algorithms performing better in most cases. The loads that are considered in this thesis are hard real-time tasks with associated deadlines. Specifically, a comparison is made between clusters with one where the head node doesn't participate in processing of the work-loads with the other where the head node does participate in processing of the work-loads. A new mathematical formula is derived for the task execution time corresponding to the new scenario of head node possessing front-end processing capability. The existing algorithms corresponding to Real-Time Divisible Load Theory are then implemented using this new formula to examine the scheduling performance in this new scenario compared to the conventional scenario where the head node lacks front-end processing capability.