This paper concentrates on the specific process of diagnosing performance problems — a process we call performance diagnosis. The more general process of performance debugging applies diagnosis and performance tuning in an iterative manner. We view performance diagnosis as a methodology (or set of methodologies) for producing an explanation of performance phenomena (in particular, performance bottlenecks) that meets a set of diagnostic requirements. It is our opinion that performance diagnosis is the key engineering issue in effective performance debugging, but that there is a general lack of support for the performance diagnosis process.

This paper discusses the Poirot approach to providing process support for performance diagnosis.We discuss the architecture of Poirot, and describe how Poirot can re-enact or rationally reconstruct a set of published performance diagnosis strategies.