Search

Performance prediction methods for component-based software systems aim at supporting design decisions of software architects during early development stages. With the increased availability of multicore processors, possible performance gains by distributing threads and processes across multiple cores should be predictable by these methods. Many existing prediction approaches model concurrent behaviour insufficiently and yield inaccurate results due to hard underlying assumptions. In this paper, we present a formal performance prediction approach for component-based systems, which is parameterisable for the number of CPUs or CPU cores. It is able to predict the response time of component services for generally distributed execution times. An initial, simple case study shows that this approach can accurately predict response times of multithreaded software components in specific cases. However, it is limited if threads change the CPU during their execution, if the effect of processor cache thrashing is present, and if the memory bus is heavily used.

Non-functional properties of software should be specified early in the development process. In a distributed process of software development, this means that quality requirements must be made explicit in the specification, and the developing party of a commissioned component needs to deliver not only the implemented component, but also a description of its non-functional properties. Based on these artefacts, a conformance check guarantees that the implemented component fulfills the performance requirements.We extend the notion of model refinement to non-functional properties of software and propose a refinement calculus for conformance checking between abstract performance descriptions of components. The calculus is based on a refinement notion that covers the performance-relevant aspects of components. The approach is applied to the Palladio Component Model as a description language for performance properties of components.