PROFPRED: A Compiler-Level IR Based Performance Prediction Framework for MPI Industrial Applications

Meng Hao, Weizhe Zhang, Laurence Tianruo Yang

Abstract: The performance prediction of parallel applications is important in high performance computing. It can discover performance bottlenecks, assist the optimization of scientific and industrial applications, and guide the job scheduling. This paper proposes PROFPRED a compiler-level IR based framework designed to predict computation and communication performance of MPI parallel industrial applications on large-scale target systems using small subsets or small-scale prototypes of target systems. We demonstrate this integrated set of techniques in our framework on Taub cluster and TianHe-2 supercomputer with seven NAS Parallel Benchmarks (NPB) and two real-world large-scale parallel applications, including CGPOP, and ASCI Sweep3D. The core algorithms of these selected benchmarks are widely used in industrial applications. The median prediction errors for these experimental applications range from 0.35% to 11.61%, and the average error is 4.28%. Compared to traditional regression-based prediction approach, PROFPRED presents better accuracy for predicting the performance of parallel applications with large scales.