TABLE OF CONTENTS We formulate a revenue/penalty-based multiresource allocation problem for multiple QoS classes. In the present context, the control policies concern the load-dispatching router and the per-server resources, although our methods and results can be applied more generally. Most previous optimal resource allocation studies of this type have focused on QoS performance guarantees based on throughput or mean response time measures. However, a critical issue for Web hosting applications and services concerns the per-request efficiency with which the differentiated services are handled, since delays experienced by clients of each service provider customer can result in lost revenue and clients for the customer. Further, such QoS performance guarantees may not be fully captured by the more standard performance metrics. To address these issues, we consider a general class of SLA models that include the tail distributions of the per-class response times in addition to standard metrics. The decision variables of the corresponding optimization problem are concerned with allocating resources to maximize the profit of hosting the customer Web sites under these SLA constraints and a general cost model.
Our problem falls within the general class of optimal resource allocation problems, but based on the foregoing nonconventional performance metrics. Recently, Menasc et al. (2000) considered the related problem of resource scheduling in Web sites with the aim of maximizing revenue. A priority scheme is proposed for scheduling requests based on the states (navigation and purchase) of the client, who will lose patience if the response time is too long. Simulations...
