Real-time scheduling of divisible loads in clusters to handle estimated execution time inaccuracies

Quality of Service (QoS) provisioning for divisible loads in clusters, enabled using real-time scheduling theory, is based on an important assumption: the scheduler knows the execution time of even' task in the workload. Information from production clusters, however, shows that user estimated task execution times are often inaccurate. In this paper, we present an approach to address this problem, where a feedback mechanism is developed to overcome the inaccuracies in user estimation of task execution time. Besides QoS provisioning of divisible loads in clusters, the proposed algorithm achieves low task reject ratio and high system utilization. Simulation evaluation shows that the new algorithm outperforms the previous approach.