Maximizing the Utility of a Computer Service using Adaptive Optimal Control

This paper describes a control system that aims to maximize the value or utility obtained from a computer service (e.g., e-commerce system and networked file system) that is shared by multiple customers. Each customer has negotiated contractual performance objectives, such as response time bounds and minimum throughput bounds, with the service provider. It is the service provider's goal to maximize the utility it gets from serving these customers workloads less any penalties for not meeting the contracted performance goals. Our design consists of a MIMO adaptive optimal controller coupled with a nonlinear optimizer. The optimizer uses the model produced by the estimator in the adaptive controller to calculate the performance goals that maximizes the utility of the service. These goals are the references to the controller that aims to meet these by dynamically allocating the share of computing resources that each customer receives. Experimental results from a 3-tier e-commerce site shows that our approach can indeed meet latency and throughput goals and maximize the utility of the system, even when the system and the workload change significantly