Cost-Efficient Construction of Performance Models

Modern high-performance applications are highly-configurable systems that provide hundreds of configuration options. Performance models offer insights into the performance of these applications and help users understand the impact of these options. Yet, crafting models for such applications proves costly due to the many configuration options and their unknown performance impacts that need to be modeled. However, some options are performance-irrelevant, and removing them can reduce construction costs without compromising accuracy. This paper explores an approach to automatically identify performance-irrelevant configuration options empirically. By leveraging established performance modeling methods, we devise cost-efficient preliminary prediction models that rely on fewer samples and analyze them to identify such options. We evaluate our approach using a real-world HPC application to demonstrate our method’s effectiveness in recognizing performance-irrelevant options and the potential to save costs for performance modeling.