Management intelligence for optimal resource allocations in network server systems

In this paper, we provide a control-theoretic treatment of the resource allocations that adaptively occur in a QoS-aware network server system. Here, the target system being controlled is a logical service point that processes the transactions requested by clients using a resource infrastructure, with a goal of maximizing the revenues. Accurate management of resource allocations with a revenue-oriented goal is quite complex, due to the interactions among various transactions that dynamically share the resources in the system (such as server nodes, disks, content caches, and network bandwidth). So, we adopt an on-line monitor-and-control approach, aided by heuristics, that iteratively adjusts the resource allocation based on the observed transaction drop rate. We undertake a case study of end-to-end QoS-adaptive data transfer to illustrate the methodology. In terms of control theory, the bandwidth allocation and the packet loss rate constitute the system input and output respectively, with the heuristics-based bandwidth adjustment strategies incorporated in a controller along the feedback loop. The use of control theory allows offering predictable convergence properties of the QoS seen by applications, while maximizing the service provider revenues.