Inconsistent system behavior causes unpredictable performance which is known to stress users; making the system perform consistently should remove this source of user stress. Operating systems currently provide the illusion that each application runs on a dedicated Virtual Machine. This paper proposes incorporating performance into this abstraction, resulting in a Virtual Private Machine. The VPM abstraction aims to improve user-perceived performance by increasing performance consistency, and it is applicable to any uservisible application, from word processors to web servers. To provide VPMs, per-resource performance models allow resources to be scheduled to meet target response times calculated for each user-visible action.

A common problem that sales consultants face in the field is the selection of an appropriate hardware and software configuration for web farms. Over-provisioning means that the tender will be expensive while under-provisioning will lead to a configuration that does not meet the customer criteria. Indy is a performance modeling environment which allows developers to create custom modeling applications. We have constructed an Indy-based application for defining web farm workloads and topologies. The paper presents an optimization framework that allows the consultant to easily find configurations that meet customers' criteria. The system searches the solution space creating possible configurations, using the web farm models to predict their performance. The optimization tool is then employed to select an optimal configuration. Rather than using a fixed algorithm, the framework provides an infrastructure for implementing multiple optimization algorithms. In this way, the appropriate algorithm can be selected to match the requirements of different types of problem. The framework incorporates a number of novel techniques, including caching results between problem runs, an XML based configuration language, and an effective method of comparing configurations. We have applied the system to a typical web farm configuration problem and results have been obtained for three standard optimization algorithms.

No abstract