Transplantation And Optimization Of Water Supply Network Simulation Calculation For "Songshan" Supercomputer System

In recent years,with the increasing proportion of pipeline transportation in my country’s energy structure,the scale of water supply network construction has gradually increased.It is very important to use the pipe network simulation technology to accurately predict the numerical state in the pipeline.However,the simulation efficiency of the commonly used pipe network simulation software is still insufficient when it is oriented to the simulation of large and complex pipe networks.High-efficiency simulation is an urgent problem that needs to be solved at present.The "Songshan" supercomputer system adopts the CPU+DCU(Deep Compute Unit)heterogeneous computing mode,which has great advantages in highly intensive computing.In order to improve the speed and performance of pipe network simulation calculation,this paper transplants the water supply pipe network simulation program to the "Songshan" platform,and continues to carry out in-depth optimization research on this heterogeneous program.It provides a more efficient parallel solution for real-time simulation analysis of water supply network.The main work and contributions of this thesis are as follows:(1)Analyze the principle and characteristics of the simulation calculation of the water supply pipe network simulation program,and realize the simulation calculation program of the CPU version.The program performance analysis technology is used to locate the computational hotspot in the simulation implementation,and the parallelism of the computationally intensive pipe element and non-pipe element in the program is analyzed and researched.(2)According to the heterogeneous parallel characteristics of "Songshan" supercomputer system,a parallel realization method of water supply pipe network simulation program is proposed.(3)According to the parallel characteristics between pipe elements and non-pipe elements,a multi-level parallel optimization method MULTPARA is proposed to solve the problem that thread blocks cannot be fully parallelized,two levels of parallelism at the coarse-grained thread-block level and fine-grained thread-level are realized for pipe elements and non-pipe elements.Aiming at the problem of thread warp differentiation existing in the calculation of some components,combined with the characteristics of massive threads of the DCU deep computing unit,a cross-thread access optimization method is proposed to realize the rational use of computing resources and improve the simulation performance.(4)For large-scale water supply network simulation calculation,a parallel simulation calculation method in multi-DCU accelerators is designed by using flow to realize component data transmission and calculation asynchronous.Aiming at the parallel computing of multi-DCU,a component partitioning scheme is proposed to reasonably divide the data of pipe elements and non-pipe elements,which solves the problem of data dependence in parallel computing of multi-DCU.In order to reduce the time-consuming of calculation and communication,the component data is transmitted to different DCU devices for calculation by stream,which realizes the overlapping of component data transmission and calculation time,and improves the calculation speed and performance of large-scale pipeline network simulation.The work done in this paper has been realized and applied on the "Songshan" supercomputer system.The simulation computing system of the water supply pipe network is the first time to realize the transplant optimization research on the "Songshan" supercomputer system.The experimental results show that the heterogeneous implementation version of the water supply pipe network simulation system is about 14.355 times faster than the CPU version,which lays a foundation for the simulation implementation of other energy pipe network simulation calculations on the "Songshan" platform.