Mining the performance of complex systems

The enormous amount of data that can be collected in any performance evaluation study of a complex system indicates the need for methodologies and systems capable of analyzing, fusing and reducing high-dimensional data spaces with very high speed. In this paper, we devise and present an adaptation of the Knowledge Discovery in Databases (KDD) framework introduced by U. Fayyad et al. (1996) that supports the above functionality for performance data of complex software/hardware system pairs. The KDD framework considered integrates database technology along with data mining techniques for uncovering patterns from performance data and static system characteristics. A case study is presented to demonstrate the effectiveness and applicability of the KDD approach for the performance evaluation of complex systems. The data mining tools utilized are general-purpose, public-domain and independent of the specific performance database involved. We are currently implementing the proposed KDD framework within an end-to-end performance evaluation system for designing complex parallel and distributed systems referred to as POEMS (Performance-Oriented End-to-end Modeling System)