Hotspot: a Framework to Support Performance Optimization on Multiprocessors

摘要

High Performance Computing (HPC) programs are usually developed by domain specialists, and they are not always experts on performance optimization. Application domain specialists rely on a systematic and unattended (and/or partially guided) method to support performance analysis and optimization. This paper presents a framework that simplifies the job, including several case studies as a validation step and as a proof of concept. The main goal is to make the data recollection task straight-forward, allowing to focus on experimentation and optimization. The framework gathers context information about the program and the underlying system, detailing scaling behavior, execution profile, resource utilization, and bottlenecks. In a nutshell, this work contributes with a performance report generator for OpenMP programs.