Cost-minimizing preemptive scheduling of mapreduce workloads on hybrid clouds

MapReduce has become the dominant programming model for processing massive amounts of data on cloud platforms. More and more enterprises are now utilizing hybrid clouds, consisting of private infrastructure owned by themselves and public clouds such as Amazon EC2, to process their spiky MapReduce workloads, which fully utilize their own on-premise resources while outsourcing the tasks only when needed. With disparate workloads of different MapReduce tasks, an efficient scheduling mechanism is in need to enable efficient utilization of the on-premise resources and to minimize the task outsourcing cost, while meeting the task completion time requirements as well. In this paper, a fine-grained model is described to characterize the scheduling of heterogeneous MapReduce workloads, and an online algorithm is proposed for joint task admission control into the private cloud, task outsourcing to the public cloud, and VM allocation to execute the admitted tasks on the private cloud, such that the time-averaged task outsourcing cost is minimized over the long run. The online algorithm features preemptive scheduling of the tasks, where a task executed partially on the on-premise infrastructure can be paused and scheduled to run later. It also achieves desirable properties such as meeting a pre-set task admission ratio and bounding the worst-case task completion time, as proven by our rigorous theoretical analysis.