A Self-adaptive HPL-Based Benchmark with Dynamic Task Parallelism for Multicore Systems

Benchmarks play a crucial role in enhancing systems, applications, and technologies across multiple domains. In scientific computing, traditional benchmarks such as HPL (High-Performance Linpack) require parameters such as the number of processes, input data size, and other execution details to reflect the characteristics of the system under test, which leads to the adoption of fine-tuning strategies. In this sense, this paper presents a novel implementation of the HPL, originally used for evaluating distributed-memory parallel architectures. The proposed benchmark, named SA-HPL, enhances HPL with self-adaptive features to facilitate the evaluation of shared-memory parallel architectures. The inclusion of novel schedulers for dynamic task parallelism, combined with simplified configuration settings, makes the SA-HPL well-suited to provide enhanced performance and efficiency without requiring fine-tuning configurations. Results show that our approach improves the throughput to 19534.50 MFLOPS and maintains an efficiency level of 74.70% compared to the baseline (with unbalanced workload) in a workstation setup, which recorded 16267.60 MFLOPS and 61.23%, respectively.