Planning Workflow Executions When Using Spot Instances In The Cloud

When running workflows in the cloud it is appealing to use spot instances that can be acquired at a fraction of the cost of on-demand instances. Unfortunately, spot instances can be revoked at any time, creating uncertainty about task completion times, which is an impairment for workflows with timeliness requirements. While workflow scheduling has been subject to extensive research, the problem of optimally scheduling deadline-constraint workflows in the cloud while dealing with the uncertainty caused by spot instance revocations has not been fully addressed. In this paper, we plan the execution of workflows in cloud environments to minimize the monetary cost while being subject to timeliness constraints. Our approach constructs a Markov Decision Process (MDP) of the workflow execution and looks up for the optimal policy taking into account the user preferences in time and cost. The optimal solution is generated offline and actions selected on-the-fly, depending on the occurrence of failures due to instance revocations. Experimental results with a real-world scientific workflow application demonstrate that, in comparison to approaches that rely on simple heuristics to schedule tasks, our planning-based approach is able to generate reliable solutions that are cheaper and able to meet deadlines.