With parallel and distributed computing entering the mainstream of computer science, it is important to ensure that parallel application codes are optimized to achieve the best possible performance. This thesis describes the design and implementation of CHITRA92, the second generation of a performance analysis system for parallel programs. CHITRA92 is unique in that it uses visualization techniques to analyze the dynamic activity of a program and produces a semi-Markov chain model of the program's behavior. This model can be parameterized to predict behavior of a program and identify the performance bottlenecks in the program. The important contributions of the CHITRA92 system are: • The dynamic activity of a program is represented through a program execution sequence (PES) and a set of program parameters. A PES description language has been defined to allow users to describe the structure of a PES state vector and instances of the state vector. • A PES is reduced to a semi-Markov chain model via a set of transformations. • Visualization is used to assist in selecting which transforms to apply to a PES. • The presence of periodic behavior in PES's can be identified using spectral analysis techniques. / Master of Science

Distributed computer systems are becoming increasingly prevalent, thanks to modern technology, and this leads to significant challenges for the software developers of these systems. In particular, in order to provide a certain service level agreement with users, the performance characteristics of the system are critical. However, developers today typically consider performance only in the later stages of development, when it may be too late to make major changes to the design. In this thesis, we propose a performance driven approach to development — based around tool support that allows developers to use performance modelling techniques, while still working at the level of program code. There are two central themes to the thesis. The first is to automatically relate performance models to program code. We define the Simple Imperative Remote Invocation Language (SIRIL), and provide a probabilistic semantics that interprets a program as a Markov chain. To make such an interpretation both computable and efficient, we develop an abstract interpretation of the semantics, from which we can derive a Performance Evaluation Process Algebra (PEPA) model of the system. This is based around abstracting the domain of variables to truncated multivariate normal measures. The second theme of the thesis is to analyse large performance models by means of compositional abstraction. We use two abstraction techniques based on aggregation of states — abstract Markov chains, and stochastic bounds — and apply both of them compositionally to PEPA models. This allows us to model check properties in the three-valued Continuous Stochastic Logic (CSL), on abstracted models. We have implemented an extension to the Eclipse plug-in for PEPA, which provides a graphical interface for specifying which states in the model to aggregate, and for performing the model checking.