- Verwendung von Continuous Integration zur Automatisierung von POP-basierten Leistungsbewertungen für HPC-Anwendungen
Schlien, Paul; Müller, Matthias S. (Thesis advisor); Wolf, Felix (Thesis advisor); Orland, Fabian (Consultant); Tomski, Felix (Consultant)
Aachen : RWTH Aachen University (2025)
Bachelor Thesis
Bachelorarbeit, RWTH Aachen University, 2025
Abstract
Performance analysis of High-Performance Computing (HPC) applications remains a highly manual and effort-intensive task, requiring expert knowledge to apply and interpret a diverse set of performance analysis tools. This slows down performance-driven software development, which is particularly critical in the HPC domain, as the manual effort required for data collection diverts time away from actual analysis and the identification of optimisation opportunities. A particularly affected use case is the Performance Optimisation and Productivity Centre of Excellence (POP CoE) project, which provides consulting services for performance analysis and optimisation for HPC applications. Hence, their services heavily rely on performance-oriented development workflows, where code and configuration changes necessitate frequent evaluation of their performance impact. To address this challenge, this thesis proposes a Continuous Integration (CI)-based workflow that automates key aspects of the POP-driven performance assessment process. A template-based framework was developed to systematise the application of performance analysis tools to HPC applications in a CI context, which has been used in the implementation of the performance assessment workflow, realised on the GitLab CI/CD platform. The focus lies on enabling automated performance data collection and laying the groundwork for continuous performance evaluation. The proposed approach was evaluated using a real-world HPC application for which a prior manual POP assessment exists. Results indicate that the template-based framework can support the automation of performance assessments under specific assumptions regarding the execution model of the application and the behaviour of performance analysis tools -- namely, non-interactivity and a predictable usage pattern. In this setup, CI serves as an orchestrator, with the primary analysis logic remaining within the performance analysis tools. Furthermore, findings indicate that more fine-grained improvements in pipeline efficiency are necessary, along with enhanced mechanisms for assessing performance trends over time. Overall, this work demonstrates the feasibility of a CI-based, template-driven performance assessment workflow and its applicability in selected HPC contexts.
Institutions
- IT Center [022000]
- Department of Computer Science [120000]
- Chair of High Performance Computing (Computer Science 12) [123010]