- Evaluation der Aussagekräftigkeit von Hardware-Performance-Countern gemessen entlang des kritischen Pfades
Brückner, Moritz; Müller, Matthias S. (Thesis advisor); Lankes, Stefan (Thesis advisor); Thärigen, Ben (Consultant)
Aachen : RWTH Aachen University (2026)
Master Thesis
Masterarbeit, RWTH Aachen University, 2026
Abstract
Critical path analysis has proven to be a useful technique for the performance analysis of parallel programs, since it is concerned with the program activities responsible for the program’s total execution time. However, if on-the-fly analysis tools are used for the analysis of critical paths, they usually cannot capture the structure of the critical paths because their capabilities are inherently limited. Other common approaches for the analysis of parallel programs make use of hardware performance counters, whose measurements are often aggregated as part of profiles or accumulated over time along threads or processes. However, this comprises all program activities independent of their impact on performance. Measurements of hardware performance counters along critical paths, as recently implemented in the On-The-Fly Critical-Path-Tool (OTF-CPT), mitigate both aforementioned problems: They only consider those program activities that determine the total execution time and may help to better understand the observed critical path even if its structure is unknown. This thesis introduces a performance analysis methodology based on output provided by OTF-CPT, which is applied to five different applications in order to evaluate the expressiveness of the tool’s measurements. The methodology compares the critical path with individual threads in order to expose its characteristics, and thereby investigates potential causes for its structure. The conducted experiments show that the measurements along the critical path can be used to determine performance factors not visible in measurements of individual threads or processes. Due to the use of hardware performance counters, it is possible to rule out certain causal relationships between hardware usage and the critical path. However, because OTF-CPT does not capture the structure of critical paths, it is not possible to map the measurement results to individual regions in the program’s source code, which constitutes a major obstacle. Nonetheless, the results show that the methodology can be used to create initial hypotheses as a basis for more in-depth analyses. This thesis also discusses the necessary conditions and implications of an adaptation of the methodology for other tools. While a direct adaptation is unlikely to work with current output capabilities of analysis tools such as Scalasca, modifying the tools themselves would allow for a successful application.
Institutions
- E.ON Energy Research Center [080052]
- Chair of High Performance Computing (Computer Science 12) [123010]
- Institute for Automation of Complex Power Systems [616310]