Performance Modeling of OpenMP Program Based on LLVM Compilation Platform

Abstract

AbstractWith the rapid development of computers and their technologies, high- performance computers (HPC) have gradually become one of the indispensable research methods in the field of basic science. However, with the development of HPC, its scale and complexity have also increased by orders of magnitude, which lead to problems such as high execution efficiency and energy consumption of parallel applications in the HPC system. So it brings many challenges to parallel programming. In order to solve the above situation, the mechanism needs to be proposed, which can predict the performance characteristics of parallel programs before they are executed. Therefore, parallel program performance prediction is the key to solving the above problems. OpenMP program is the most common shared memory parallel program, so it is an important part of the research of parallel program performance evaluation. This article uses the LLVM compilation platform to convert the OpenMP source code into LLVM intermediate code (IR) through Clang, statically and dynamically analyzes the IR, and then obtains the OpenMP performance model to predict the OpenMP execution time and optimal execution time. Based on the modeling of OpenMP parallel applications, we can get the performance model of the OpenMP program. Based on this model, we can predict the performance and scalability of OpenMP-related applications in HPC systems, find program and system bottlenecks, and guide program performance optimization.KeywordsHPCOpenMP parallel programLLVM compilation platformStatic analysisDynamic analysisPerformance modelingScalability prediction