Exploiting Performance Scaling and Predictive Modeling for SA-HPL: From Multicore to Distributed Systems
Cassiano Rista, Marcelo Teixeira, Mauro Fonseca · IEEE Access · 2025
High-Performance Computing (HPC) systems rely on computational benchmarks to evaluate and compare the performance of scientific applications across diverse architectures. However, the manual configuration required by traditional benchmarks, such as High-Performance Linpack (HPL), often leads to laborious and error-prone tuning processes. To overcome this, we introduce a self-adaptive HPL version (SA-HPL), now significantly enhanced with a Generalized Stochastic Petri Nets (GSPN)-based predictive modeling layer for autonomous parameter tuning. Building upon our foundational work focused on multicore systems, this study extends SA-HPL to support distributed environments through a hybrid MPI+OpenMP architecture, facilitating dynamic task parallelism and self-reconfiguration. A primary contribution of this research is the extensive experimental validation of SA-HPL, demonstrating its adaptability and scalability across diverse execution contexts, from multicore setups to distributed clusters, encompassing up to 32 processing cores. The integrated GSPN model efficiently simulates various workload scenarios, guiding performance-aware reconfigurations without the need for exhaustive empirical tuning. Our results confirm exceptional predictive accuracy in cluster environments, with Mean Absolute Percentage Error (MAPE) values as low as 0.12% for throughput and 0.03% for efficiency. These findings underscore the effectiveness of our approach in providing scalable, accurate, and cost-efficient performance evaluation and planning for high-performance computing systems.