Computer system for accelerating fluid simulation through cooperation of DPU and CPU

Abstract

The invention discloses a computer system for accelerating fluid simulation through cooperation of a DPU (Data Processing Unit) and a CPU (Central Processing Unit), which comprises one or more than two computing nodes, the computing nodes are communicated through a network, each computing node is provided with the same number of CPUs and DPUs, each computing node is provided with one or more than two CPUs, and the plurality of CPUs are communicated through internal connection. The CPU is used for being responsible for iterative solution of a fluid simulation control equation and writing an iterative result into the memory; each computing node is provided with one or more than two DPUs (Data Processing Unit), the DPUs directly access a controller by adopting a memory, directly read memory data without passing through a CPU (Central Processing Unit) memory controller, calculate the residual error of an iteration result and / or verify whether the convergence criterion is met, and when the current iteration result meets the convergence criterion, the CPU completes the on-going iteration solution, so that the convergence criterion is met. And the DPU performs residual calculation on the iteration result, performs post-processing and outputs the result, and transfers residual calculation and post-processing to the DPU, so that the utilization rate of a CPU (Central Processing Unit) is improved, and the simulation precision and efficiency are further improved.

Description

Technical Field

[0001] The present invention relates to the field of large-scale parallel solutions in computational fluid dynamics, and in particular to a computer system in which a DPU and a CPU collaborate to accelerate fluid simulation. Background Art

[0002] Fluid simulation is a technology that simulates fluid flow through computers. It is widely used in aerospace, automotive, energy and other fields. In fluid simulation, the CPU undertakes the core computing tasks. Its multi-core processing power and multi-threading characteristics are crucial to handling large-scale parallel computing tasks, especially in fluid dynamics problems involving a large number of iterations and complex calculations. However, with the continuous improvement of the complexity and accuracy requirements of simulation models, the demand for computing resources has increased dramatically. With the continuous improvement of the complexity and accuracy requirements of simulation models, the demand for computing resources has increased dramatically. However, the improvement of CPU performance is restricted by the lithography process, and the development speed has slowed down, making it difficult for the CPU to meet the needs of high-performance fluid simulation.

[0003] The performance bottlenecks in fluid simulation are mainly reflected in two aspects: one is the floating-point computing performance of the CPU, which is crucial for performing numerical calculations; the other is the memory read and write speed, because fluid simulation needs to frequently read and write data from the memory. For example, matrix iterative solutions are usually limited by the floating-point performance of the CPU, while residual calculations of physical fields such as pressure and flow rate are more limited by the memory read and write speed. Different hardware bottlenecks of different algorithms give rise to heterogeneous algorithms with different computing units.

[0004] In the prior art, there are computer systems for accelerating fluid simulation, such as the patent publication number CN118714020B, a computer network system and a computer system. This patent improves the communication bottleneck between CPUs by optimizing the network topology, thereby improving the efficiency of parallel computing, but there is a bottleneck in the memory read and write speed.

[0005] Another example is the announcement number CN118245039B, a parallel algorithm transplantation optimization method based on domestic accelerators. Through the heterogeneity of CPU and GPU, a parallel algorithm is developed to accelerate fluid simulation, and the floating-point computing performance of GPU is used to improve the parallel computing speed. The overall computing performance is significantly improved, but it also faces some shortcomings, such as high hardware cost, CPU parallel algorithm duplication, and difficulty in transplanting to GPU.

[0006] The above two different types of existing technologies do not solve the bottleneck problem of memory reading and writing speed in fluid simulation, and have limited scope of application. Summary of the invention

[0007] In order to address the deficiencies in the prior art, the present invention aims to provide a computer system in which a DPU and a CPU collaborate to accelerate fluid simulation, thereby increasing memory access bandwidth, reducing CPU core waiting time, and thereby improving fluid simulation efficiency.

[0008] To achieve the above objectives, a computer system for accelerating fluid simulation by cooperating between DPU and CPU is proposed, which includes one or more computing nodes. The computing nodes communicate with each other through a network, and each computing node is configured with the same number of CPUs and DPUs. Each computing node has one or more CPUs, and multiple CPUs communicate with each other through internal connections. The CPU is responsible for iteratively solving the fluid simulation control equations and writing the iterative results into the memory; Each computing node has one or more DPUs. The DPUs use a direct memory access controller to directly read memory data without going through the CPU memory controller, calculate the residual of the iteration result and / or verify whether it meets the convergence criterion. When the current iteration result meets the convergence criterion, the CPU completes the ongoing iterative solution, and the DPU calculates the residual of the iteration result, performs post-processing and outputs the result.

[0009] Furthermore, the computer system uses the DPU to collaborate with the CPU to accelerate the solution of the fluid simulation algorithm. The specific steps are as follows: S1: The computational domain is allocated to each computing node. The computational domain grid and physical quantities are stored in the memory. The CPU in each computing node sets the initial conditions and boundary conditions of the computational domain. S2: Each computing node performs parallel computation of fluid simulation according to task allocation. Specifically, The CPU advances the time step based on the initial conditions or the calculation results of the previous time step, iteratively solves the physical quantity in each time step and stores it in the memory. The DPU reads the boundary conditions and iterative results in the memory to calculate the residual and judge whether the residual meets the convergence criterion; the CPU reads the convergence judgment result of the DPU, if not, the CPU continues to iterate the solution; if satisfied, the CPU completes the ongoing iterative solution, and the DPU calculates the residual of the iterative result; S3: DPU performs post-processing and result output, and determines whether this time step is the last time step. If not, it enters the next time step, repeating S2 and S3 in each time step until the last time step. DPU writes the solved physical quantity to the hard disk.

[0010] Furthermore, the specific steps in step S1 include: S11: Divide the computational domain into several blocks with relatively balanced numbers of grids according to the grid density and geometric characteristics of the computational domain; S12: Marking several units of the block interface so that the calculation results of the block nodes of different calculation nodes are equal during iterative calculation; S13: configure a parallel computing environment so that each block can be processed independently and in parallel on different computing nodes, and each computing node can access the physical quantities of the computing domain in other computing nodes through the network; S14: Specify initial values ​​for each physical quantity in the computational domain, set boundary conditions, and apply the boundary conditions to the boundary grid of the computational domain.

[0011] Furthermore, the time step is advanced based on the initial value or the calculation result of the previous time part. The specific steps of parallel calculation of fluid simulation based on physical quantity calculation nodes in each time step are as follows: ①. The CPU iteratively solves the momentum conservation equation and updates the flow velocity Ui. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the equation N+1 times, obtains the N+1 iteration result, and updates the flow velocity Ui for subsequent steps. At the same time, the DPU calculates the residual of the N+1 iteration result as the final residual record. Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ② The CPU iteratively solves the pressure Poisson equation and updates the pressure p. The DPU calculates the residual of the pressure Poisson equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the N+1th time, obtains the N+1th iteration result and updates the pressure p for subsequent steps. At the same time, the DPU calculates the residual of the N+1th iteration result as the final residual record. Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ③. Use pressure p to correct flow rate Ui. DPU calculates the residual of mass conservation equation and makes convergence judgment. If convergence judgment is not met, return to step ① and continue iterative calculation until convergence criterion is met or the maximum number of inner loops between step ① and step ③ is reached; ④. Solve other physical quantities by using the corresponding equations in step ①. At the same time, use the solved physical quantities to solve the unsolved physical quantities until all physical quantities are solved; ⑤. After all physical quantities are solved, the DPU calculates the residual of the continuity equation through the solved physical quantities and determines whether the residual meets the convergence criterion of the outer loop between steps ① and ⑤; if not, return to step ① to continue iterative calculation until the convergence criterion of the outer loop between steps ① and ⑤ is met or the maximum number of outer loops between steps ① and ⑤ is reached, and then enter step S3.

[0012] Furthermore, the specific steps of post-processing and result output include: After reaching the convergence criterion, the DPU is used to perform auxiliary calculations such as simulation calculation result monitoring and real-time post-processing; Through DPU direct memory access technology, a data write request is directly initiated to the CPU memory controller, and after completion of the processing, the result is written to the memory; Or use PCIe pass-through technology to write the calculation results and post-processing results to the hard disk directly or after compression.

[0013] Furthermore, in step ④, other physical quantities include the turbulent viscosity coefficient vt, the turbulent kinetic energy k and the specific dissipation rate ω, and the specific solution steps include: [1] The CPU iteratively solves the turbulent kinetic energy k equation and updates the turbulent kinetic energy k. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the equation N+1 times, obtains the N+1th iteration result, and updates the turbulent kinetic energy k for subsequent steps. Or the CPU reaches the maximum number of iterations, obtain the iteration result of the maximum number of iterations and update the turbulent kinetic energy k for subsequent steps; [2] The CPU calculates the specific dissipation rate ω through the turbulent kinetic energy k, and then calculates the turbulent viscosity coefficient vt.

[0014] Technical effects and advantages of the present invention: 1. The present invention effectively reduces the waiting time of the CPU core by offloading the residual calculation task in the fluid simulation to the DPU, thereby accelerating the fluid simulation calculation; 2. The asynchronous parallel method allows the CPU to continue iterating while the DPU is calculating the residual. After the residual meets the convergence condition, the current iteration is completed and the calculation result is retained. According to experience, adding one iteration can further reduce the residual and improve the accuracy.

[0015] 3. The present invention uses DMA technology to enable the DPU to directly initiate a data write request to the CPU memory controller, unloading the CPU's result output task. This process bypasses the CPU's intervention, allowing computing resources and storage resources to be transmitted independently, significantly reducing the CPU's I / O burden and data transmission delay, thereby accelerating fluid simulation.

[0016] 4. The DPU uses a PCIe slot to connect to the computer's motherboard and is compatible with current common servers, making it easy to upgrade existing servers. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the computer system principle; Figure 2 A schematic diagram of the main process of the method for accelerating fluid simulation for a computer system; Figure 3 is a hardware schematic diagram of a computer system embodiment; Figure 4 Schematic diagram of the PIMPLE algorithm flow for accelerating fluid simulation in computer systems; Figure numerals: 1, CPU; 2, CPU core; 3, CPU memory controller; 4, PCIe controller; 5, memory; 6, hard disk; 7, DPU; 8, direct memory access controller; 9, network controller; 10, computing node; 11, DAC high-speed cable. DETAILED DESCRIPTION

[0018] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0019] Refer to Figure 1 As shown, the present invention provides a computer system for accelerating fluid simulation by collaboration between a DPU and a CPU, including one or more computing nodes, the computing nodes communicate with each other through a network, each computing node is configured with the same number of CPUs and DPUs, each computing node has one or more CPUs, and multiple CPUs communicate with each other through internal connections. The CPU is responsible for iterative solution of fluid simulation control equations and writes the iterative results into a memory; each computing node has one or more DPUs, and the DPUs all use a memory direct access controller to directly read memory data without passing through the CPU memory controller, calculate the residual of the iterative result and / or verify whether it meets the convergence criterion. When the current iterative result meets the convergence criterion, the CPU completes the ongoing iterative solution, and the DPU calculates the residual of the iterative result, performs post-processing and outputs the result.

[0020] For specific embodiments, refer to Figure 3As shown in the figure, the computer system includes two ASUS RS720A-E11-RS12 rack servers, each of which is called a computing node. Each computing node contains two AMD 7773X CPUs, and the CPUs integrate memory controllers and PCIe controllers; each CPU is connected to a DPU and a network controller through a PCIe controller; the motherboard of each computing node integrates a DMA controller and is connected to a hard disk; each CPU is connected to 12 memory sticks; the network controllers on the two computing nodes are connected in pairs through DAC high-speed cables.

[0021] Table 1 Hardware selection

[0022] Refer to Figure 2 and Figure 4 As shown, on the computer system, a method for fluid simulation is provided in which a DPU and a CPU collaborate to accelerate the PIMPLE algorithm. The method is based on the numerical solution of incompressible flow using the PIMPLE algorithm. The numerical solution of incompressible flow is an important topic in fluid mechanics, and its key algorithm is pressure-velocity coupling, including SIMPLE, PISO, and PIMPLE. Among them, the PIMPLE method combines the iterative correction of SIMPLE with the multi-step pressure correction of PISO, and is suitable for solving steady-state and unsteady-state incompressible flows, with a wide range of adaptability and high solution accuracy. Based on the hardware system of the embodiment, the specific steps are as follows: S1: The computational domain is assigned to two computing nodes. The computational domain grid and physical quantities are stored in the memory of the two computing nodes respectively. The CPU sets the initial conditions and boundary conditions of the computational domain. Specifically: According to the grid density and the geometric characteristics of the computational domain, the computational domain is divided into 48 blocks. In this embodiment, the number of blocks is equal to the number of memories, and the number of grids in the blocks is relatively balanced, so that the computational load of each computing node is balanced. The units on the block interface are marked to ensure that the calculation results are equal when the block interface is iterated at different computing nodes.

[0023] Configure the parallel computing environment so that each block can be processed independently and in parallel on different computing nodes, and each computing node can access the physical quantities of the computing domain in other computing nodes through the network; specify the initial value for each physical quantity in the computing domain, and set the corresponding boundary conditions. The boundary conditions are assigned to the boundary grid using fixed values ​​or gradient values.

[0024] S2: Two computing nodes perform parallel calculations of fluid simulation according to task allocation. The CPU starts from the initial conditions or the calculation results of the previous time section and iteratively solves the physical quantities in the calculation domain. The physical quantities in this embodiment include flow velocity Ui, pressure p, turbulent viscosity coefficient vt, turbulent kinetic energy k and specific dissipation rate ω. The above physical quantities are coupled. This application uses the PIMPLE algorithm to decouple and predict, solve and correct the physical quantities in turn. After each iteration is completed, the iteration results are written to the memory. The DPU reads the boundary conditions in the memory and the iteration results of the physical quantities in the calculation domain, calculates the residual of the physical quantity, and determines whether the residual meets the convergence criterion. In this embodiment, the convergence criterion is set to 10 -6 When the DPU calculates the residual, the CPU continues to perform iterative solution and writes the iterative result into the memory; the CPU reads the convergence judgment result of the DPU, and if it has not converged, it continues to perform iterative solution, and if it has converged, it goes to step S3.

[0025] Specifically: the time step is advanced based on the initial value or the calculation result of the previous time part, and the fluid simulation parallel calculation is performed based on the physical quantity calculation node in each time step: ①. The CPU performs iterative solution of the momentum conservation equation and updates the flow velocity Ui. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU performs N+1th iterative solution, obtains the N+1th iterative result and updates the flow velocity Ui for subsequent steps. At the same time, the DPU calculates the residual of the N+1th iterative result as the final residual record. For example: the CPU performs the first iterative solution, and then the CPU performs the second iterative solution. At the same time, the DPU calculates the residual of the first iterative solution for convergence judgment. After the DPU judges that the residual of the result of the first iteration satisfies the convergence, the CPU completes the second iteration and uses the result of the second iteration as the final result for subsequent steps. At the same time, the DPU calculates the residual of the second iterative solution and directly uses it as the final residual record without performing a convergence judgment. Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ② The CPU iteratively solves the pressure Poisson equation and updates the pressure p The DPU calculates the residual of the pressure Poisson equation and makes a convergence judgment, repeating the cycle until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU performs N+1 iterations to obtain the N+1 iteration results and updates the pressure. p For subsequent steps, the DPU calculates the residual of the N+1 iteration results as the final residual record; Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the updated flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ③. Use the pressure p to correct the flow rate Ui. The DPU calculates the residual of the mass conservation equation and makes a convergence judgment. If the convergence judgment is not satisfied, return to step ① and continue the iterative calculation until the convergence criterion is satisfied or the maximum number of inner loops between steps ① and ③ is reached (set to 3 times in this embodiment); ④. Solve other physical quantities by using the corresponding equations to solve step ①. At the same time, solve the unsolved physical quantities by using the solved physical quantities until all physical quantities are solved. In step ④, other physical quantities include turbulent viscosity coefficient vt, turbulent kinetic energy k and specific dissipation rate ω. The specific solution steps include: [1] The CPU iteratively solves the turbulent kinetic energy k equation and updates the turbulent kinetic energy k. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the equation N+1 times, obtains the N+1th iteration result, and updates the turbulent kinetic energy k for subsequent steps. Or the CPU reaches the maximum number of iterations, obtain the iteration result of the maximum number of iterations and update the turbulent kinetic energy k for subsequent steps; [2] The CPU calculates the specific dissipation rate ω through the turbulent kinetic energy k, and then calculates the turbulent viscosity coefficient vt.

[0026] ⑤. After all physical quantities are solved, the DPU calculates the residual of the continuity equation through the solved physical quantities and determines whether the residual meets the convergence criterion of the outer loop between steps ① and ⑤; if not, return to step ① to continue iterative calculation until the convergence criterion of the outer loop between steps ① and ⑤ is met or the maximum number of outer loops between steps ① and ⑤ is reached (set to 20 times in this embodiment), and then enter step S3.

[0027] S3: DPU performs post-processing and result output, and determines whether this time step is the last time step. If not, it enters the next time step, repeating S2 and S3 in each time step until the last time step, when DPU writes the solved physical quantity to the hard disk. The specific steps of post-processing and result output are: After reaching the convergence criteria, the DPU is used to perform auxiliary calculations such as simulation calculation result monitoring and real-time post-processing. Through the DPU direct memory access technology, a data write request is directly initiated to the CPU memory controller, and after the post-processing is completed, the result is written to the memory. Alternatively, the PCIe pass-through technology is used to write the calculation results and post-processing results to the hard disk directly or after compression.

[0028] Specifically, in this embodiment, the DPU outputs the numerical calculation results of the flow field such as the velocity field U, the pressure field p, and the turbulent kinetic energy k, and draws contour maps of the flow field, pressure distribution, turbulence intensity, etc. to evaluate the fluid behavior and calculation accuracy; S2 and S3 are repeated until the last time step is completed. After the last time step is completed, the DPU compresses the physical quantity and writes it to the hard disk.

[0029] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A computer system for accelerating fluid simulation by cooperating between a DPU and a CPU, characterized in that: It includes one or more computing nodes, which communicate with each other through the network. Each computing node is configured with the same number of CPUs and DPUs. Each computing node has one or more CPUs, and multiple CPUs communicate with each other through internal connections. The CPU is responsible for iteratively solving the fluid simulation control equations and writing the iterative results into the memory; Each computing node has one or more DPUs. The DPUs use a direct memory access controller to directly read memory data without going through the CPU memory controller, calculate the residual of the iteration result and / or verify whether it meets the convergence criterion. When the current iteration result meets the convergence criterion, the CPU completes the ongoing iterative solution, and the DPU calculates the residual of the iteration result, performs post-processing and outputs the result.

2. The computer system for accelerating fluid simulation by cooperation between DPU and CPU according to claim 1, characterized in that: The computer system uses DPU and CPU to accelerate the solution of fluid simulation algorithm. The specific steps are as follows: S1: The computational domain is allocated to each computing node. The computational domain grid and physical quantities are stored in the memory. The CPU in each computing node sets the initial conditions and boundary conditions of the computational domain. S2: Each computing node performs parallel computation of fluid simulation according to task allocation. Specifically: The CPU advances the time step based on the initial conditions or the calculation results of the previous time step, iteratively solves the physical quantity in each time step and stores it in the memory. The DPU reads the boundary conditions and iterative results in the memory to calculate the residual and judge whether the residual meets the convergence criterion; the CPU reads the convergence judgment result of the DPU, if not, the CPU continues to iterate the solution; if satisfied, the CPU completes the ongoing iterative solution, and the DPU calculates the residual of the iterative result; S3: DPU performs post-processing and result output, and determines whether this time step is the last time step. If not, it enters the next time step, repeating S2 and S3 in each time step until the last time step. DPU writes the solved physical quantity to the hard disk.

3. The computer system for accelerating fluid simulation by cooperation between DPU and CPU according to claim 2, characterized in that: The specific steps in step S1 include: S11: Divide the computational domain into several blocks with relatively balanced numbers of grids according to the grid density and geometric characteristics of the computational domain; S12: Marking several units of the block interface so that the calculation results of the block nodes of different calculation nodes are equal during iterative calculation; S13: configure a parallel computing environment so that each block can be processed independently and in parallel on different computing nodes, and each computing node can access the physical quantities of the computing domain in other computing nodes through the network; S14: Specify initial values ​​for each physical quantity in the computational domain, set boundary conditions, and apply the boundary conditions to the boundary grid of the computational domain.

4. The computer system for accelerating fluid simulation by cooperation between DPU and CPU according to claim 2, characterized in that: The time step is advanced based on the initial value or the calculation result of the previous time part. The specific steps for parallel calculation of fluid simulation based on physical quantity calculation nodes in each time step are: ①. The CPU iteratively solves the momentum conservation equation and updates the flow velocity Ui. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the equation N+1 times, obtains the N+1 iteration result, and updates the flow velocity Ui for subsequent steps. At the same time, the DPU calculates the residual of the N+1 iteration result as the final residual record. Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ② The CPU iteratively solves the pressure Poisson equation and updates the pressure p. The DPU calculates the residual of the pressure Poisson equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the N+1th time, obtains the N+1th iteration result and updates the pressure p for subsequent steps. At the same time, the DPU calculates the residual of the N+1th iteration result as the final residual record. Or the residual does not satisfy the convergence judgment and reaches the maximum number of iterations, the iteration result of the maximum number of iterations is obtained and the flow rate Ui is updated for subsequent steps. At the same time, the DPU calculates the residual of the iteration result of the maximum number of iterations as the final residual record; ③. Use pressure p to correct flow rate Ui. DPU calculates the residual of mass conservation equation and makes convergence judgment. If convergence judgment is not met, return to step ① and continue iterative calculation until convergence criterion is met or the maximum number of inner loops between step ① and step ③ is reached; ④. Solve other physical quantities by using the corresponding equations in step ①. At the same time, use the solved physical quantities to solve the unsolved physical quantities until all physical quantities are solved; ⑤. After all physical quantities are solved, the DPU calculates the residual of the continuity equation through the solved physical quantities and determines whether the residual meets the convergence criterion of the outer loop between steps ① and ⑤; if not, return to step ① to continue iterative calculation until the convergence criterion of the outer loop between steps ① and ⑤ is met or the maximum number of outer loops between steps ① and ⑤ is reached, and then enter step S3.

5. The computer system for accelerating fluid simulation by cooperation between DPU and CPU according to claim 2, characterized in that: The specific steps of post-processing and result output include: After reaching the convergence criterion, the DPU is used to perform auxiliary calculations such as simulation calculation result monitoring and real-time post-processing; Through DPU direct memory access technology, a data write request is directly initiated to the CPU memory controller, and after completion of the processing, the result is written to the memory; Or use PCIe pass-through technology to write the calculation results and post-processing results to the hard disk directly or after compression.

6. The computer system for accelerating fluid simulation by cooperation between DPU and CPU according to claim 4, characterized in that: In step ④, other physical quantities include the turbulent viscosity coefficient vt, the turbulent kinetic energy k and the specific dissipation rate ω. The specific solution steps include: [1] The CPU iteratively solves the turbulent kinetic energy k equation and updates the turbulent kinetic energy k. The DPU calculates the residual of the momentum conservation equation and makes a convergence judgment. This cycle repeats until the Nth residual calculated by the DPU satisfies the convergence judgment. The CPU iterates and solves the equation N+1 times, obtains the N+1th iteration result, and updates the turbulent kinetic energy k for subsequent steps. Or the CPU reaches the maximum number of iterations, obtain the iteration result of the maximum number of iterations and update the turbulent kinetic energy k for subsequent steps; [2] The CPU calculates the specific dissipation rate ω through the turbulent kinetic energy k, and then calculates the turbulent viscosity coefficient vt.

Images