Dynamic auto-scaling and scheduling of deadline constrained service workloads on IaaS clouds

Cloud systems are becoming attractive for many companies. Rather than over-provisioning the privately owned infrastructure for peak demands, some of the work can be overspilled to external infrastructure to meet deadlines. In this paper, we investigate how to dynamically and automatically provision resources on the private and external clouds such that the number of workloads meeting their deadline is maximized. We specifically focus on jobs consisting of multiple interdependent tasks with a priori an unknown structure and even adaptable at runtime. The proposed approach is model -driven: knowledge on the job structure on the one hand; and resource needs and scaling behavior on the other hand. Information is built up based on monitoring information and simulated 'what-if-scenarios. Using this dynamically constructed job resource model, the resources needed by each job in order to meet its deadline is derived. Different algorithms are evaluated on how the required resources and jobs are scheduled over time on the available infrastructure. The evaluation is carried out using synthetic workloads. (C) 2016 Elsevier Inc. All rights reserved.