Performance data mining: Automated diagnosis, adaption, and optimization

The development of performance measurement and analysis techniques and tools for high-performance parallel and distributed systems has made it possible to capture a wealth of data about application and system performance behavior. These data embody the effects of interacting, performance factors found in the program, its algorithms, the architecture and hardware, and the system software, whose interdependent performance relationships grow ever more complex as the computing environment increases in sophistication. Nevertheless, the user is still, for the most part, placed in the central decision-making role in the use of the techniques/tools, the interpretation of the resulting performance information, and the guidance for program or system modification. Recent work has sought to move human decisionmaking out of the performance measurement–diagnosis–optimization loop by employing “intelligent” methods based on automated performance measurement, knowledge-based diagnosis frameworks, online, adaptive performance control, and predictive performance models built from detailed empirical analysis. The term “performance data mining” is used to characterize this work. Already back in 1997 in association with the International Conference on Supercomputing, we organized a workshop on performance data mining for high-performance parallel and distributed systems. To the best of our knowledge, this was the first scientific event dedicated to the exploitation of data mining techniques in the field of performance evaluation. The papers enclosed in this volume have their roots in this pioneering workshop. The topics discussed reflect different perspectives on the performance data mining research question and on the focus for effective strategies and techniques. In the following sections, we first discuss our views on the concept of performance data mining. We then introduce each paper and comment on how we see the ideas presented contributing to the problem understanding and solution. We conclude with thoughts for future research.