Performance modeling for MPI applications with low overhead fine-grained profiling

摘要

MPI applications have been widely used in the scientific computing and cloud computing fields. Understanding how these applications will scale on HPC and cloud platforms is essential for users and system designers. However, achieving this task is difficult because of the complexity of applications and systems. In this work, we propose an automatic, fine-grained profiling approach based on linear regression. Different from those in previous studies, our approach profiles MPI applications at the basic block level. Using this fine-grained profiling level, we can provide users with detailed information on how each part of the application will scale on hundreds or thousands of cores. We can also determine the scalability limit. Additionally we use two methods to reduce the profiling cost to less than 50% of the runtime of the original application. We test our approach on TianHe-2, which is ranked number 2 on the Top500 list as of November 2017, and Taub clusters, which is developed by UIUC. The median prediction errors of our approach are 8% and 13% for two NPB benchmarks and two real applications, respectively. We also compare our approach with PEMOGEN. The results show that our approach is more accurate on large process counts. (C) 2018 Elsevier B.V. All rights reserved.