Novel techniques and methods for performance measurement, analysis and monitoring of cluster and grid applications

Abstract

The evolution of parallel and distributed architectures and programming paradigms for performance-oriented program development challenges the state of technology for performance tools. Coupling different programming paradigms such as message passing and shared memory programming for hybrid cluster computing is one example for high demands on performance analysis tools that are capable to cope with applications for multiple programming models and target architectures. This dissertation presents a classification of temporal overheads for parallel programs that can be used to explain sources of performance problems. A highly customizable instrumentation and measurement system