OpenMP/MPI hybrid programming-based communication optimization method of CFD simulation program

Abstract

The invention discloses an OpenMP / MPI hybrid programming-based communication optimization method of a CFD simulation program. The method specifically includes the following steps: carrying out area decomposition on a calculation domain according to communication loads and calculation loads to divide the same into multiple sub-domains on which parallel solving can be carried out; determining network topology relationships among all grid blocks according to setting situations of {processes and thread groups}, and creating and initializing MPI persistence communication relationships among the processes; after calculation advancing which is independent among threads ends, activating MPI channels among the processes for message communication, using shared memory at the same time to communicatein the thread groups to realize boundary data exchange among the grid blocks; and calculating residuals in the thread groups, collecting a maximum residual among the processes, if the maximum residualstill does not meet a convergence condition, returning to the grid advancing step for iteration, and otherwise, ending parallel calculation. Therefore, parallel efficiency of the CFD simulation program can be effectively improved.

Description

technical field [0001] The invention belongs to the intersection field of computational fluid dynamics and computer, and in particular relates to a communication optimization method of a CFD simulation program based on OpenMP / MPI hybrid programming. Background technique [0002] Computational Fluid Dynamics (CFD for short) is an interdisciplinary subject accompanied by the rapid rise of computers. It provides cheap simulation, design and optimization tools for the theory and engineering research of fluid mechanics, and widely serves traditional fluid mechanics and fluid engineering fields such as aviation, aerospace, ships, water conservancy, etc. The essence of the CFD method is to discretize the governing equations of the flow field to a series of grid nodes by computational mathematics to obtain their discrete numerical solutions. Therefore, the scale of grid points determines the complexity of solving CFD problems. At the beginning of computer development, it was mainl...

AI Technical Summary

Problems solved by technology

In order to simulate computational fluid dynamics with the help of a multi-core supercomputer, the CFD program is also parallelized, and the additional communication consumption between processes / threads in the parallel process has also become the bottleneck of the entire simulation program.

Supercomputing is mostly based on the Symmetrical Multi-Processing (SMP) architecture, which has the characteristics of multiple cabinets in the whole machine, nodes in the cabinet, multiple CPUs in the node, and multi-core CPUs. It uses a single shared memory to communicate with OpenMP or message Neither transfer communication MPI can make full use of the characteristics of supercomputing itself

[0004] At present, the use of hybrid programming in CFD simulation programs is relatively extensive. In order to simplify the communication process by utilizing the multi-core characteristics of supercomputing, most simulation programs mostly adopt a pure MPI architecture. When the program scale approaches 10,000 cores, frequent message communication between processes The resulting mesh bandwidth pressure creates an obvious bottleneck

A small number of programs that use OpenMP / MPI mixed programming only stay in OpenMP's fine-grained acceleration of local for loops, lack of corresponding optimization strategies, and do not take advantage of coarse-grained OpenMP shared memory communication

Images