- Bewertung und Vergleich von Datenplatzierungs-Optimierungswerkzeugen für heterogene Speicherarchitekturen
Huckebrink, Ben-Jay; Müller, Matthias S. (Thesis advisor); Lankes, Stefan (Thesis advisor); Klinkenberg, Jannis (Consultant)
Aachen : RWTH Aachen University (2025)
Bachelor Thesis
Bachelorarbeit, RWTH Aachen University, 2025
Abstract
The memory-related demands of scientific applications rise at an ever-accelerating pace. However, traditional dynamic random access memory (DRAM) has not kept up with these increasing memory capacity, speed, and energy efficiency demands. In response, heterogeneous memory systems employing multiple memory types, such as non-volatile memory (NVM) or high-bandwidth memory (HBM), alongside DRAM have risen to prevalence. Leveraging the advantages of such systems involves placing individual application data structures into different memory types depending on their memory access behaviors. Since manually conducting such a placement optimization requires detailed application knowledge and a large time investment, previous research developed data placement optimization frameworks to automate this process and improve the placement decisions made. However, previous research on these frameworks has not adequately evaluated their efficacy. Most existing work tests only the execution time performance of the frameworks' placement decisions, leaving the frameworks' user experience and energy efficiency benefits unquantified. Crucially, existing research also does not compare the different frameworks against one another. In combination, these shortcomings impede research on future frameworks, since the specific strengths and weaknesses of already existing approaches remain unknown, meaning their weaknesses cannot be improved systematically. In this thesis, I address this shortage by evaluating and comparing three state-of-the-art data placement optimization frameworks in-depth. For this purpose, I develop a custom, highly configurable synthetic benchmark that can systematically alter its memory access behaviors. This configurability allows me to detail specific strengths and weaknesses of each framework's placement optimization algorithm and quantify their impact in terms of the execution time and energy efficiency the made placement decisions achieve. By also testing the frameworks on four proxy applications, I assess the real-world implications of the identified advantages and disadvantages. Further, using the proxy applications, I uncover shortcomings in the frameworks' user experience. Based on my observations, I propose modifications to the frameworks to improve their decision-making and their user experience.
Institutions
- IT Center [022000]
- Faculty of Computer Science [120000]
- Chair of High Performance Computing (Computer Science 12) [123010]