Co-design at system and component level: examples from the DEEP and EPI projects
Optimizing the configuration of HPC systems to meet user requirements requires in-depth knowledge of application profiles and hardware limitations. A quantitative methodology, founded on the development of application-based benchmarks executed on representative hardware platforms, simulators, and models, proves invaluable for this objective. Benchmarking sheds light on how application codes and their core components perform on HPC systems, enabling the identification of performance bottlenecks and opportunities for enhancement on the software side. Moreover, it aids in understanding the impact of specific hardware features on application performance, whether positive or negative. During this presentation, we will share our experiences with benchmark-driven co-design approaches, derived from two European projects, namely DEEP and EPI. In DEEP, our focus was primarily on the system level, while in EPI the target were on processor and core-level aspects. We will describe the differences between both approaches and the challenges that we found.