M.: The need for a whole-system view of performance

Abstract. Middleware usually runs on top of a powerful execution platform. Often that platform includes extensive runtime libraries, a virtual machine, an operating system, and modern complex hardware. The un-derstanding of middleware performance is complicated by intricate interactions between the application, the middleware, and the components of the underlying platform. The functional interfaces of the components of such a platform are clearly defined. They range from the API specification, over the virtual machine specification, the system call interface, to the processor’s instruction set architecture. The functionality of a system component can be understood by just looking at that compo-nent’s functional specification. The performance of a system component, on the other hand, can often not be understood in isolation. This is because performance interactions between components can be very intricate, non-linear, and involve unpredictable delays. We propose temporal vertical profiling [HSDH04] as an approach for whole system performance understand-ing. Temporal vertical profiling captures the temporal system behavior on all layers, and of all components, of the system. A trace of values for performance metrics from system components is captured and analyzed to identify correlations between such metrics. Correlation is the first step in the identification of causal releation-ships between metrics and components. And these causal relationships are the fundamental building blocks for performance understanding