Research On The Key Technologies Of The OpenFOAM Oriented Performance Analysis

With the development of computing technology, the performance of parallel computer systems is increasing day by day. However, when it comes to the practical use of HPC in the problem solving, the performance gap between peak performance and real performance is becoming more and more obviously as the expansion of the simulation scale, which is called the Programming Wall Problem. On the one hand, it has always been a difficult issue to develop parallel programs; On the other hand, the performance analysis of massively parallel applications is much more complex than serial ones. The development pattern based on parallel framework can reduce the difficulties of development by hiding some implementation details, but it is still a question worth studying to improve the performance of programs based on this pattern. Take the massively parallel CFD development as an example, this paper launched a research revolves around the performance analysis of parallel CFD development framework.First of all, this paper puts forward a multi-user oriented performance analysis method,Mu PAM, for the CFD parallel development framework. The core idea of the method is the performance analysis for multi-user, especially for those non-computer professional users. After the classification of users participating in the CFD simulations, the different performance analysis requirements of different users are analysed, and an overall performance analysis framework is designed.Then, we design a specific performance analysis tool called Mu PAM-OF in OpenFOAM based on the framework of Mu PAM. The platform features of Open FOAM is researched and analysed, and the structure framework of Mu PAM-OF is presented, which has three functional modules:the performance monitoring module, the performance analysis module and visualization module. The design of the performance data sets are also given, and some classification and sorting has been carried out from the angle of optimization and implementation.Next, we implement the Mu PAM-OF after going into its key technologies on the basis of the given framework. The study and implementation of each module has been launched respectively. As for the monitoring module, a dynamic performance detection method is used to collect performance data, and some optimizations have been taken to improve the collection efficiency such as a custom performance data file format, etc; The performance analysis module analysed and quantified the various performance factors from the aspect of system and algorithm with the method of static performance analysis; While the visualization module is implemented based on the combining of VTK and Para View.Finally, some experiments have been conducted in order to validate the practicality and validity of Mu PAM-OF. The experimental results show that the Mu PAM-OF can not only monitor and display the performance of CFD simulations successfully, but also can locate performance bottlenecks, which has practical significance for the improvement and optimization of large-scale parallel CFD simulations.