Advanced Memory Data Structures for Scalable Event Trace Analysis

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