Author Description

This dissertation presents a novel approach for the analysis of event traces from parallel program runs. Event tracing is an established method for the analysis and optimization of performance and scalability of parallel and High Performance Computing (HPC) applications. The Complete Call Graph (CCG) data structure is introduced as an alternative in-memory representation for huge event traces. It exploits frequently found redundancies and provides customized data reduction in a combination of lossless and lossy compression. It achieves data compression by factors of ten to factors of hundred and requires no explicit decompression during the analysis. Furthermore, adapted evaluation procedures allow to reduce the computational effort in the same order of magnitude as the data compression. The thesis includes a comprehensive discussion of the state-of-the-art of event tracing and related work, a detailed presentation of the design of the CCG data structure, an elaborate description of the algorithms for construction, compression, and analysis of CCGs, and an extensive experimental validation of all components.