Performance monitoring and analysis environment for distributed memory MIMD programs

This thesis studies event monitoring techniques that are used for collecting, filtering and visualising event traces from parallel programs. Implementations of two experimental monitoring systems are presented. The first system is a hybrid implementation which uses extra hardware to collect event traces. The second system is a software implementation which was implemented on the Edinburgh Concurrent Supercomputer. These two systems can gather event traces from the parallel programs at a very low cost. The event abstraction mechanism is used for filtering these event traces. The generic and applicationspecific performance metrics are achieved by using event abstraction techniques to replace event patterns with new abstract events which are used for visualising performance related behaviour. The strengths and weaknesses of the event abstraction approach are discussed in the context of performance analysis and visualisation of message passing parallel programs.