- Profiling-Basierte und Statistische Erkenntnisse von On-the-Fly Analyse des kritischen Pfades
Beckmann, Lorenz; Müller, Matthias S. (Thesis advisor); Lankes, Stefan (Thesis advisor); Thärigen, Ben (Consultant)
Aachen : RWTH Aachen University (2026)
Bachelor Thesis
Bachelorarbeit, RWTH Aachen University, 2026
Abstract
The increasing complexity of modern high-performance computing (HPC) applications, driven by parallel programming models such as MPI and OpenMP, makes performance analysis increasingly challenging. Identifying bottlenecks like load imbalances or synchronisation overheads remains difficult, as interactions between parallel processes and threads grow more complex. Traditional trace-based analysis tools provide detailed insights but often require substantial post-processing effort. This thesis addresses these challenges by extending the On-The-Fly Critical-Path Tool (OTF-CPT), a lightweight performance analysis tool developed at RWTH Aachen University. The primary objective of this work is to enhance the analytical depth of the OTF-CPT while maintaining its low-overhead, on-the-fly nature. To achieve this, all data previously collected but not exposed by the tool is now stored in CSV files. Based on this dataset, new statistical summaries were implemented to enable detailed visualisation of performance data distributions. These extensions allow for a more precise interpretation of application behaviour compared to the original tool's focus on mean and maximum values. Furthermore, an existing profiling component was integrated, extending the computation of performance model factors from a single global set to all individual parallel regions, thereby enabling direct attribution of inefficiencies to the source code. The extended tool was evaluated using the NEST simulator on the CLAIX-2023 cluster. The integrated profiling and statistical capabilities provide helpful insights into the execution behaviour, revealing a parallel region with high load imbalance responsible for a slowdown of the overall execution. The new capabilities were compared to the established post-mortem toolset Score-P and Cube. While Cube offers fine-grained, interactive insights for all regions, the extended OTF-CPT provides similar features and additional metrics such as useful execution time while maintaining its on-the-fly nature. Although performance model factors are now computed for all regions, statistical summaries remain limited to the complete execution rather than being available per region as in Cube. To assess runtime overhead, the application was executed using different analysis tools. Results show that the runtime increase in the extended OTF-CPT is mainly due to the finalisation phase of the profiling component rather than interference during execution, ensuring metrics accurately reflect the original behaviour.
Institutions
- IT Center [022000]
- Chair of High Performance Computing (Computer Science 12) [123010]