A high-efficiency analysis method for limit bearing of arrow body thin-wall structure heterogeneous acceleration

Abstract

The application discloses a kind of arrow body thin wall structure limit bearing heterostructure acceleration high-efficiency analysis method, comprising: the grid division of arrow body thin wall structure image model is obtained with several grid units and nodes;According to several grid units, search and division are carried out to image model, and several unit combinations are obtained;According to GPU load balancing strategy, unit combination is distributed to GPU computing unit;According to central difference method, node, unit and unit combination are calculated dynamically using GPU computing unit, and calculation result is obtained;Using asynchronous transmission and double buffering method, calculation result is output to external CPU.The application improves the calculation efficiency, while maintaining the accuracy of numerical simulation.

Description

Technical Field

[0001] This invention relates to an accelerated and efficient analysis method for the ultimate load-bearing capacity of thin-walled rocket body structures, belonging to the field of high-performance numerical simulation analysis technology. Background Technology

[0002] Thin-walled structures in rocket bodies, as critical components of the main load-bearing structure of launch vehicles, often experience instability due to axial compression loads caused by launch overloads. Explicit dynamic simulations of quasi-static processes are frequently used to solve the ultimate load-bearing capacity of such structures, addressing non-convergence issues caused by strong nonlinearities such as large deformations, material nonlinearities, and complex contacts. However, existing explicit dynamic analysis methods face challenges such as high computational resource consumption and long computation times, making them difficult to adapt to complex and ever-changing design requirements. Commercial software like Abaqus' explicit dynamics is limited by computational efficiency, insufficient for handling large-scale models, and unable to meet the demands of rapid product iteration. There is an urgent need to utilize high-performance computing resources such as GPUs and FPGAs to improve analysis speed. Summary of the Invention

[0003] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a high-efficiency and accelerated analysis method for the ultimate load-bearing capacity of thin-walled rocket bodies, thereby accelerating the explicit dynamic analysis process.

[0004] The technical solution of this invention is:

[0005] This invention discloses an accelerated and efficient method for the ultimate load-bearing heterogeneous analysis of thin-walled rocket body structures, comprising:

[0006] The image model of the thin-walled structure of the rocket body is meshed to obtain several mesh elements and nodes;

[0007] Based on several grid units, the image model is searched and divided to obtain several unit combinations;

[0008] According to the GPU load balancing strategy, the unit combination is allocated to the GPU computing unit;

[0009] Based on the central difference method, the GPU computing unit is used to perform dynamic calculations on the nodes, units, and unit combinations to obtain the calculation results.

[0010] The calculation results are output to an external CPU using asynchronous transmission and double buffering.

[0011] Furthermore, in the above method, the specific method for searching and dividing the image model is as follows:

[0012] The combinations of units are divided into combinations a, b, c, and d; where combination a consists of 2*2*2 units; combination b consists of 2*2 units; combination c consists of 2*1 units; and combination d consists of one unit.

[0013] Assign priorities A, B, C, and D to combinations a, b, c, and d respectively, where A > B > C > D.

[0014] Image models are searched and divided according to priority from highest to lowest to obtain several unit combinations.

[0015] Furthermore, in the above method, the GPU load balancing strategy is specifically as follows:

[0016] The units of type combination a, combination b, combination c and combination d are distributed equally to several GPUs in sequence.

[0017] Furthermore, in the above method, the GPU computing unit performs dynamic calculations on the nodes, units, and unit combinations, specifically as follows:

[0018] S40, Initialize time step and computation time;

[0019] S41, Calculation velocity and t at time n+1 Displacement at any moment;

[0020] S42, Calculate t n+1 Inner strength at all times external force and the next time step;

[0021] S43, according to t n+1 Inner strength at all times external force And the next time step, calculate t n+1 acceleration at time and t n+1 Speed ​​in a given moment;

[0022] S44. Accumulate the calculation time; determine whether the calculation time exceeds the threshold. If yes, output the displacement, velocity, acceleration and internal force of the node and exit the calculation; otherwise, proceed to step S45.

[0023] S45. Repeat steps S41 to S44 according to the next time step.

[0024] Furthermore, in the above method, the specific method for calculating the next time step is as follows:

[0025] t n+1 =t n +Δt n+1

[0026]

[0027] Where, Δt n+1 It is the time step, t n and t n+1 This corresponds to the time when the nth step is executed. It is the time interval between the nth and n+1th moments.

[0028] Furthermore, in the above method, the calculation velocity and t at time n+1 The displacement at any given time is as follows:

[0029]

[0030] in, for The speed of time For t n The speed of time For t n Acceleration at any moment; The midpoint of the time step interval; For t n Displacement at any given moment.

[0031] Furthermore, in the above method, the calculation of t n+1 acceleration at time and t n+1 The speed at any given moment, specifically:

[0032]

[0033] in, Let c be the velocity and acceleration at time n+1. i For damping, These are the values ​​of external force and internal force, respectively. for Speed ​​in time.

[0034] The advantages of this invention over the prior art are as follows:

[0035] (1) This invention utilizes GPU heterogeneous parallel technology to achieve efficient acceleration of the explicit dynamic analysis process, thereby solving the problem that the explicit dynamic algorithm implemented on a single processor in the existing technology is difficult to meet the computational efficiency requirements of the ultimate bearing capacity of thin-walled structures.

[0036] (2) This invention proposes a heterogeneous accelerated and efficient analysis method for the ultimate load-bearing capacity of thin-walled rocket bodies. It makes full use of the many-core characteristics of GPUs to perform calculations on such large-scale independent tasks using GPU threads, which significantly improves the efficiency of explicit dynamics calculations while maintaining the accuracy of numerical simulations. This solves the problems of high computational resource consumption and long calculation time faced by existing explicit dynamics methods in analyzing thin-walled rocket bodies. Attached Figure Description

[0037] Figure 1 This is a diagram illustrating the overall acceleration and optimization framework of the explicit simulation parallel algorithm of this invention.

[0038] Figure 2 This is the execution flowchart of the explicit simulation algorithm for the structure of this invention;

[0039] Figure 3 This is the preprocessing partitioning diagram of the present invention; (a) is the combination of eight units, (b) is the combination of four units, (c) is the combination of two units, and (d) is the combination of one unit.

[0040] Figure 4 This is a diagram of the multi-GPU load balancing strategy of this invention;

[0041] Figure 5 This is a schematic diagram of asynchronous transmission and double buffering in this invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] This invention discloses an accelerated and efficient method for the ultimate load-bearing heterogeneous analysis of thin-walled rocket body structures, comprising:

[0044] The image model of the thin-walled structure of the rocket body is meshed to obtain several mesh elements and nodes;

[0045] Based on several grid units, the image model is searched and divided to obtain several unit combinations;

[0046] According to the GPU load balancing strategy, the unit groups are allocated to the GPU computing units;

[0047] Based on the central difference method, the GPU computing unit is used to perform dynamic calculations on nodes, units, and unit combinations to obtain the calculation results.

[0048] The calculation results are output to an external CPU using asynchronous transmission and double buffering.

[0049] Preferably, the image model is searched and divided, and the specific method is as follows:

[0050] The combinations of units are divided into combinations a, b, c, and d; where combination a consists of 2*2*2 units; combination b consists of 2*2 units; combination c consists of 2*1 units; and combination d consists of one unit.

[0051] Assign priorities A, B, C, and D to combinations a, b, c, and d respectively, where A > B > C > D.

[0052] Image models are searched and partitioned according to priority from highest to lowest, resulting in several unit combinations. Preferably, the GPU load balancing strategy is as follows:

[0053] The units of type combination a, combination b, combination c and combination d are distributed equally to several GPUs in sequence.

[0054] Preferably, such as Figure 2 As shown, the GPU computing unit performs dynamic calculations on nodes, cells, and combinations of cells, using the following method:

[0055] S40, Initialize time step and computation time;

[0056] S41, Calculation velocity and t at time n+1 Displacement at any moment;

[0057] S42, Calculate t n+1 Inner strength at all times external force and the next time step;

[0058] S43, according to t n+1 Inner strength at all times external force And the next time step, calculate t n+1 acceleration at time and t n+1 Speed ​​in a given moment;

[0059] S44. Accumulate the calculation time; determine whether the calculation time exceeds the threshold. If yes, output the displacement, velocity, acceleration and internal force of the node and exit the calculation; otherwise, proceed to step S45.

[0060] S45. Repeat steps S41 to S44 according to the next time step.

[0061] Preferably, the next time step is calculated using the following method:

[0062] t n+1 =t n +Δt n+1

[0063]

[0064] Where, Δt n+1 It is the time step, t n and t n+1 This corresponds to the time when the nth step is executed. It is the time interval between the nth and n+1th moments.

[0065] Preferably, calculation velocity and t at time n+1 The displacement at any given time is as follows:

[0066]

[0067] in, for The speed of time For t n The speed of time For t n Acceleration at any moment; The midpoint of the time step interval; For t n Displacement at any given moment.

[0068] Preferably, t is calculated. n+1 acceleration at time and t n+1 The speed at any given moment, specifically:

[0069]

[0070] in, Let c be the velocity and acceleration at time n+1. i For damping, These are the values ​​of external force and internal force, respectively. for Speed ​​in time.

[0071] Example

[0072] This invention provides an explicit dynamic heterogeneous computing framework based on GPU parallelism, such as... Figure 1 As shown. Under the CUDA architecture, a computation method of one CUDA thread corresponding to one computation block is adopted. Based on the GPU storage structure, the memory access method and computation process are optimized to complete the solution process of the explicit structural simulation algorithm.

[0073] Design a unit preprocessing algorithm for parallel computing to improve reusability and reduce memory access overhead by optimizing unit combination and allocation. The algorithm allocates units into four basic cases based on node reuse, such as... Figure 3As shown, elements with high node reusability are preferentially grouped and assigned to the same parallel thread block. Taking a first-order hexahedral mesh in 3D space as an example, calculating the internal force and next time step size of each element requires reading the data of eight nodes on that element. However, some of the data required by adjacent elements is duplicated, and two adjacent elements will share the data of four nodes on the same face. In the CUDA storage structure, threads in the same parallel thread block can share data through shared memory. Therefore, the mesh elements processed in a block can be placed as adjacent as possible to share more nodes, reduce the number of nodes required to compute these mesh elements, and thus reduce memory access overhead. In order to allocate appropriate mesh elements to each parallel thread block and make the elements computed in a parallel thread block share as many nodes as possible, data preprocessing is required before parallel computation. Since the relative positions between elements do not change during the deformation of the simulated object, only one preprocessing is needed before all parallel computations, and subsequent parallel operations can be computed according to the partitioning scheme obtained by the preprocessing. For this preprocessing algorithm, in order to reduce its algorithm complexity while still completing the partitioning task well. Design a greedy-based preprocessing algorithm for cell partitioning. Specifically, this preprocessing algorithm divides cell combinations into four basic combinations based on node reuse, such as... Figure 3 As shown. Combination a consists of 2*2*2 units, requiring an average of only 27 / 8 = 3.375 nodes per unit, making it the combination with the highest node reuse. Combination b consists of 2*2 units, requiring an average of 18 / 4 = 4.5 nodes per unit, making it the combination with the second highest node reuse. Combination c consists of 2*1 units, requiring an average of 12 / 2 = 6 nodes per unit. Combination d has only one unit, requiring an average of 8 nodes per unit.

[0074] Based on a multi-GPU load balancing strategy, in the preprocessing scheme for a multi-GPU environment, since the units processed by each thread block are as adjacent as possible, and adjacent thread blocks share more units than other thread blocks, the threads in the block can be scheduled to different GPU devices evenly according to the allocation order. This ensures that the finite element units processed by each GPU are as adjacent as possible, concentrated in a certain region, while the finite element units processed by different GPUs are as far apart as possible, thereby achieving a multi-GPU task load balancing scheduling strategy. Figure 4 As shown.

[0075] A data transmission strategy based on asynchronous transmission and double buffering allows other computational tasks to be executed simultaneously with data transmission. Double buffering utilizes two buffers used alternately; during data transmission, one buffer is used for computation, and the other for data transmission. Once computation is complete, the two buffers switch roles to continue the next round of computation and data transmission. Double buffering not only enables better concurrent execution of computation and data transmission tasks but also effectively solves the problem of competition between computation and data transmission. While the previous round of computation is still in progress, using another buffer for data transmission avoids data consistency issues, further ensuring the correctness of asynchronously transmitted data. Furthermore, double buffering improves system robustness and reduces computational failures caused by data transmission errors. In summary, the combined application of data compression, asynchronous transmission, and double buffering significantly reduces data transmission overhead between the GPU and CPU, improves computational efficiency, and ensures data consistency and correctness, thereby enabling more efficient distributed computing in large-scale explicit finite element calculations, such as… Figure 5 As shown.

[0076] To address the low efficiency of existing explicit dynamic analysis methods based on the central difference method, a parallel strategy based on GPUs (Graphics Processing Units) can be constructed to solve the technical problem of low efficiency in explicit dynamic simulation. The details are as follows:

[0077] Please see Figure 1 , Figure 1 The GPU-parallel explicit dynamics computation framework provided for this application example comprises the following steps:

[0078] S101, Preprocess the input model and prepare the data.

[0079] S102, Construction of an explicit dynamic parallel computing framework based on the central difference method.

[0080] Step 1: Preprocess the input model:

[0081] Taking a three-dimensional first-order hexahedral mesh element as an example, in parallel computing at the element level, each thread needs to obtain the physical quantities of 8 nodes of the element, calculate the internal force of the element, and the next time step. However, some nodes of adjacent elements are repeated. In the CUDA storage structure of this application example, threads in the same parallel thread block share data through shared memory. This element preprocessing algorithm divides adjacent elements into the same parallel thread block, allowing the elements being processed to share more nodes, thereby reducing memory access overhead and improving execution efficiency. The execution flow of the preprocessing algorithm is described as follows:

[0082] Before parallel computing, the model data is preprocessed, and the units are combined and allocated according to the node reuse situation. Figure 3 Four basic cases, for each combination of units, the resulting divisions Figure 3 (a) The more (a) there are, the higher the reusability of the unit's combined nodes; the partitioned Figure 3 (d) The more elements there are, the lower the reusability of the unit's combined nodes. To maximize node reusability, when the preprocessing algorithm partitions the current parallel block into units, it prioritizes checking if there are any adjacent, unpartitioned nodes of the form […] at the boundary of the units already allocated to the parallel thread block. Figure 3 If a combination of units in (a) exists, add it to the computation scope of the parallel thread block and repeat this step; otherwise, check if there is a combination of units in the form of (a). Figure 3 (b) The unit combination, and so on, until the unit division of the parallel thread block is completed.

[0083] Based on the aforementioned unit preprocessing algorithm, a multi-GPU load balancing strategy is implemented in the preprocessing stage. In this multi-GPU load balancing strategy, for a multi-GPU environment, the preprocessing scheme ensures that the units processed by each thread block are as adjacent as possible, and adjacent thread blocks share more units than other thread blocks. Therefore, threads within a block can be scheduled to different GPU devices sequentially and evenly according to their allocation order. This ensures that the finite element units processed by each GPU are as adjacent as possible, concentrated in a specific region, while the finite element units processed by different GPUs are as far apart as possible, thereby achieving a multi-GPU task load balancing scheduling strategy. Figure 5 As shown.

[0084] Step 2: Construction of an explicit dynamic heterogeneous computational framework based on the central difference method:

[0085] When performing explicit dynamic calculations based on the central difference method, in each time step iteration, the velocity at the next half of the time step, the displacement at the next time step, the force at the current time step, and the velocity and acceleration at the next time step are updated sequentially. Therefore, we design methods to calculate the displacement, velocity, and acceleration of nodes using nodes as the granularity kernel function, and to calculate the overall force (internal force) of the element and the next time step using the element as the granularity kernel function.

[0086] 1) Initialization: Divide the solution domain into finite element parts and calculate based on the initial boundary conditions. and

[0087] 2) Update time:

[0088] t n+1 =t n +Δt n+1

[0089]

[0090] 3) Calculation velocity and t at time n+1 Displacement at any moment

[0091]

[0092] 4) Calculate t n+1 Inner strength at all times external force and the next time step

[0093] 5) Calculate t n+1 acceleration at time and t n+1 Time speed:

[0094]

[0095] 6) If t n+1 =T, meaning the current time is the end of the solution time interval, then end the loop; otherwise, jump back to step 2) to continue the iterative calculation.

[0096] In a single time-step calculation iteration, node-level calculations (displacement and velocity) are performed first, followed by element-level calculations (internal forces and calculations for the next time step), and finally, node-level calculations (velocity and acceleration) are performed again. From this perspective, if kernel functions are designed with nodes and elements as parallel granularities separately, each time-step calculation iteration requires calling the kernel function three times. However, when there is no data transfer between the device and host sides between two time-step calculation iterations, the node-level calculations for the previous and next time steps can be combined into a single kernel function. Therefore, each time-step calculation iteration can reduce the overhead of starting the kernel function by one.

[0097] In the parallel kernel function with units as the parallel granularity, each parallel thread block reads the node data corresponding to the unit partitioning scheme obtained from the preprocessing algorithm into the shared memory unique to that parallel block. Then, each thread in the parallel block retrieves the node data corresponding to that unit from the shared memory according to the sequence number and performs calculations.

[0098] After completing a specific round of computational iterations, intermediate process data needs to be output. Therefore, the data is unloaded from GPU memory to CPU host memory. (See [link / reference]). Figure 5 Asynchronous transfer and double buffering are used. Asynchronous transfer allows other tasks to be performed simultaneously with data transmission without waiting for the transmission to complete. Double buffering alternates between two buffers, one for computation and the other for data transmission. Once the computation is complete, the two buffers switch roles to continue the next round of computation and data transmission.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0100] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

[0101] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for analyzing the limit bearing heterogeneous acceleration high efficiency of a thin-walled structure of an arrow body, characterized in that, The method comprises the following steps: Grid division is performed on the image model of the thin-walled structure of the arrow body, to obtain a plurality of grid units and nodes; According to the plurality of grid units, searching and dividing the image model to obtain a plurality of unit combinations; According to the GPU load balancing strategy, the unit combinations are distributed to the GPU calculation units; According to the central difference method, the nodes, units and unit combinations are calculated by the GPU calculation units to obtain calculation results; The calculation results are output to the external CPU by using the asynchronous transmission and double buffering method; The searching and dividing of the image model are performed in the following manner: The combinations of units are divided into combinations a, b, c, and d; among them, combination a consists of 2 2 Composed of 2 units; combination b consists of 2 Composed of 2 units; combination c consists of 2 Combination d consists of one unit; combination d consists of one unit. The combination a, the combination b, the combination c and the combination d are respectively assigned with the priority A, B, C and D, wherein A > B > C > D; The searching and dividing of the image model are performed in the following manner: The GPU load balancing strategy is as follows: The unit combinations of the types of the combination a, the combination b, the combination c and the combination d are evenly distributed to the plurality of GPUs.

2. The method of claim 1, wherein the method is a high-efficiency analysis method for the limit load of the thin-walled structure of an arrow body, characterized in that, The GPU calculation units calculate the nodes, units and unit combinations in the following manner: S40, initializing the time step and the calculation time; S41, calculating Instantaneous speed and Instantaneous displacement; S42, calculate internal forces external forces and the next time step; S43、According to internal forces , external forces and the next time step, calculate instantaneous accelerations and instantaneous velocities; S44, accumulating the calculation time; judging whether the calculation time exceeds the threshold value, if yes, outputting the displacement, velocity, acceleration and internal force of the node, and exiting the calculation; if no, entering step S45; S45, repeating steps S41-S44 according to the next time step.

3. The method of claim 2, wherein the method is a high-efficiency analysis method for the limit bearing of the thin-wall structure of the arrow body, characterized in that, The next time step is calculated in the following manner: wherein, is a time step, and is the corresponding time at the running of the nth step, is the first and n+1 intermediate time.

4. The method of claim 2, wherein the method is a high-efficiency analysis method for the limit bearing of the thin-wall structure of the arrow body, characterized in that, The calculation Instantaneous speed and Instantaneous displacement, in particular: wherein, is the velocity at the time instant, is the velocity at the time instant, is the acceleration at the time instant; is the time step interval middle time instant; is the displacement at the time instant.

5. The method of claim 2, wherein the method is a high-efficiency analysis method for the limit bearing of the thin-wall structure of the arrow body, characterized in that: The calculation Instantaneous acceleration and Instantaneous velocity, in particular: wherein, are the velocity and acceleration at time n + 1, respectively, is the damping, , are the external and internal force values, respectively, is the velocity at time n.

Images