Multiobjective cloud capacity planning for time-varying customer demand

Service providers who dynamically scale cloud resources can significantly lower costs while providing a service level that conforms to a service level agreement. To do this, service providers must understand the tradeoffs within provisioning algorithms such as utilization versus system availability or the impact of service level agreement timescale on utilization. Within the context of three provisioning algorithms from existing literature, we analyze the tradeoff of service availability versus utilization using a two-dimensional Pareto analysis. We also analyze the impact on utilization and system availability of using hourly, daily, weekly, or yearlong timescales as the basis of service level agreement. We evaluate model performance using historical data of the Virtual Computing Laboratory (VCL), a cloud computing environment at North Carolina State University. We show that a simple heuristic planning model, whereby a fixed reserve capacity is maintained, provides better service availability and utilization performance than other models for all service level timescales. The fixed reserve capacity model is also shown to be Pareto optimal.