An automated profiling subsystem for QoS-aware services

The advent of QoS-sensitive Internet applications such as multimedia and the proliferation of priced performance-critical applications such as online trading mise a need for building server systems with guaranteed performance. The new applications must run on different heterogeneous platforms, provide soft performance guarantees commensurate with platform capacity: and adapt efficiently to upgrades in platforms resources over the system's lifetime. Profiling the application for the purposes of providing QoS guarantees on each new platform becomes a significant under: taking. Automated profiling mechanisms must be built to enable efficient computing of QoS guarantees tailored to platform capacity and facilitate wide deployment of soft performance-guaranteed systems on heterogeneous platforms. In this paper we investigate the design of the automated profiling subsystem; an essential component of future "general-purpose" QoS-sensitive systems. The subsystem estimates application resource requirements and adapts the software transparently to the resource capacity of the underlying platform. A novel aspect of the proposed profiling subsystem is its use of estimation theory for profiling. Resource requirements are estimated by correlating applied workload with online resource utilization measurements. We focus explicitly on profiling server software. The convergence and accuracy of our online profiling techniques are evaluated in the context of an Apache web server serving both static web pages and dynamic content. Results show the viability of using estimation theory for automated online profiling and for achieving QoS guarantees.