Cost-aware scheduling of deadline-constrained task workflows in public cloud environments

Public cloud computing infrastructure offers resources on-demand, and makes it possible to develop applications that elastically scale when demand changes. This capacity can be used to schedule highly parallellizable task workflows, where individual tasks consist of many small steps. By dynamically scaling the number of virtual machines used, based on varying resource requirements of different steps, lower costs can be achieved, and workflows that would previously have been infeasible can be executed. In this paper, we describe how task workflows consisting of large numbers of distributable steps can be provisioned on public cloud infrastructure in a cost-efficient way, taking into account workflow deadlines. We formally define the problem, and describe an ILP-based algorithm and two heuristic algorithms to solve it. We simulate how the three algorithms perform when scheduling these task workflows on public cloud infrastructure, using the various instance types of the Amazon EC2 cloud, and we evaluate the achieved cost and execution speed of the three algorithms using two different task workflows based on a document processing application.