A model-based framework for autonomous and self-adaptive benchmarking in HPC systems

Abstract

High-Performance Computing (HPC) benchmarks are essential for evaluating the efficiency ofscientific applications across different architectures. However, traditional benchmarks like the High-Performance Linpack (HPL) require extensive manual configuration and tuning, making the process time-consuming, error-prone, and difficult to reproduce. This PhD thesis introduces the Self-Adaptive HPL (SA-HPL), a novel-generation benchmark that incorporates self-adaptation mechanisms based on principles of autonomic computing. The SA-HPL is built on three main pillars: (i) the development of self-adaptive modules that eliminate the need for manual tuning, (ii) its implementation on multicore systems with support for various dynamic scheduling policies, and (iii) an extension to distributed environments via a hybrid MPI+OpenMP architecture, which enhances the benchmark’s portability and autonomy. In addition, the SA-HPL is integrated with a predictive model based on Generalized Stochastic Petri Nets (GSPN), which allows for interpretable estimation of throughput and efficiency. This integration accelerates the exploration of configuration alternatives without the need for exhaustive execution runs. Extensive experiments on multicore environments and distributed clusters with up to 32 cores confirm the approach’s precision and scalability, achieving a Mean Absolute Percentage Error (MAPE) as low as 0.12% for throughput and 0.03% for efficiency. By consolidating autonomous tuning, predictive simulation, and hybrid parallelism, this PhD thesis establishes the SA-HPL as a portable and architecture-independent benchmark. This contributes to adaptive benchmarking methodologies that can support more efficient performance evaluations and system planning in current and future HPC environments.