Parallel computing method for cluster dynamics approach in DCU-CPU hybrid architecture

Abstract

This invention provides a parallel computing method for cluster dynamics based on a DCU-CPU hybrid architecture, belonging to the field of nuclear material irradiation simulation technology. The method includes: S1, establishing data structure information on the DCU to store the data required for computation; S2, transferring the data required for computation from the CPU to the DCU; S3, allocating the solution task of the established mesh region to the DCU; S4, on the CPU, calling the solver, establishing the initial values ​​and equation matrix for the system of equations, and solving the system of equations; S5, transmitting the solution results to the DCU; S6, the DCU, based on the received solution results and the received computational data, calculating the initial values ​​and equation matrix required for the next time step; S7, the DCU transmitting the calculation results back to the CPU; S8, repeating S4-S7 until the required computation time for simulation is reached. Using this invention, the spatiotemporal scale of cluster dynamics simulation methods can be expanded, and the computation time of cluster dynamics simulation can be reduced.

Description

Technical Field

[0001] This invention relates to the field of nuclear material irradiation simulation technology, and in particular to a parallel computing method for cluster dynamics methods oriented towards a DCU-CPU hybrid architecture. Background Technology

[0002] Cluster dynamics (CD) is a numerical simulation method developed based on mean-field rate theory. This method ignores the spatial information and interaction forces of atoms, focusing instead on the size and concentration of clusters formed by atomic aggregation. It can simulate evolutionary processes over medium to long periods and belongs to the mesoscopic scale simulation method. Because cluster dynamics can provide relatively accurate model descriptions and simulations of the microscopic dynamic evolution behavior of materials, it is currently applied in many related fields of material microscopic evolution mechanism research. Taking nuclear reactors as an example, the service performance of their internal materials largely depends on the dynamic behavior of material defects generated under irradiation conditions (such as irradiation damage to structural materials and irradiation swelling caused by nuclear fuel fission gases). Therefore, using cluster dynamics to study and understand the microscopic evolution mechanism of material defects is crucial to ensuring the safe operation of nuclear reactors, improving service performance, and extending service life. In nuclear reactor research and applications, cluster dynamics has successfully simulated microscopic processes such as cluster precipitation in pressure vessels and fission gas migration.

[0003] With the expansion of applications, traditional cluster dynamics methods have revealed their bottlenecks in computation and modeling. Computationally, since cluster dynamics methods rely on solving a set of partial differential equations, the number of equations in the system typically exceeds 10. 6 Furthermore, the equations are mostly rigid, so solving the cluster dynamics equations requires an enormous amount of computation, which is difficult to accomplish with the current computing power of CPUs (central processing units). From a modeling perspective, because traditional cluster dynamics simulations do not consider the spatial correlation between clusters, and clusters are assumed to be in a homogeneous medium, the cluster dynamics model is no longer applicable when simulating rapidly changing environmental information or cluster migration processes.

[0004] Space-dependent cluster dynamics simulation methods are an improved approach to cluster dynamics models developed in recent years. By incorporating diffusion term information and spatially partitioning the model, these methods can overcome the aforementioned shortcomings in certain scenarios. However, this approach increases the computational burden of solving the cluster dynamics equations, leading to excessively long computation times and making some simulations impossible due to the computational demands, thus limiting the applicability of this method.

[0005] The Dawning supercomputer, based on a traditional CPU, incorporates a Deep Computing Unit (DCU), forming a heterogeneous hybrid architecture. The DCU is based on a general-purpose computing on graphics processing units (GPGPU) architecture, compatible with general-purpose CUDA-like environments, and suitable for program design and porting in computationally intensive scientific computing and other fields. It can effectively improve program computational efficiency and shorten program execution time. However, due to its relatively complex architecture compared to a single CPU, program design based on actual computational models presents certain challenges for designers. Summary of the Invention

[0006] This invention provides a parallel computing method for cluster dynamics simulations based on DCU-CPU hybrid architectures. This method expands the spatiotemporal scale of cluster dynamics simulations and reduces the computation time. The technical solution is as follows:

[0007] This invention provides a parallel computing method for cluster dynamics in a DCU-CPU hybrid architecture, comprising:

[0008] S1, establish data structure information on the DCU to store the data required for computation; wherein, the data structure includes: a data structure for basic cluster information, a management structure for cluster information, a management structure for grid information, and a data structure for basic cluster reaction information;

[0009] S2, the CPU transmits the data required for computation to the DCU;

[0010] S3, assign the task of solving the established grid region to the DCU terminal;

[0011] S4, on the CPU side, calls the solver to establish the initial values ​​and equation matrix for solving the system of equations, and then solves the system of equations.

[0012] S5 transmits the solution results of the system of equations to the DCU.

[0013] S6, the DCU calculates the initial values ​​and equation matrix required for the next time step based on the received solution results and the data required for the calculation.

[0014] S7, the DCU end transmits the calculation results back to the CPU end;

[0015] S8, repeat S4-S7 until the required computation time for the simulation is reached.

[0016] Furthermore, the data structure of the basic information of the cluster includes one or more of the following: size, number, diffusion coefficient, binding energy, migration energy, reaction radius, and concentration of a single type of cluster.

[0017] Furthermore, the cluster information management structure includes one or more of the following: lattice size of the simulated object, defect atom size, reaction information storage address, number of reactions, and cluster information storage address; wherein, the cluster information storage address points to a data structure of basic cluster information, and the reaction information storage address points to a data structure of basic cluster reaction information.

[0018] Furthermore, the management structure for the grid information includes one or more of the following: the number of grids, the size of the grids, and the temperature of the grids.

[0019] Furthermore, the data structure for the basic information of the cluster reaction includes one or more of the following: reaction type, storage address of cluster information participating in the reaction, and reaction rate coefficient; wherein the storage address of cluster information participating in the reaction points to the data structure for the basic information of the cluster.

[0020] Furthermore, the assignment of the established mesh region solution task to the DCU includes:

[0021] The solution domain is decomposed by meshing, and the computational tasks on a single mesh are used as computational units and allocated to threads on the DCU, forming a task partitioning method of mesh-thread and region-thread pool.

[0022] Furthermore, the DCU terminal calculates the initial values ​​and equation matrix required for the next time step based on the received solution results and the received data required for the calculation, including:

[0023] Based on the received solution results, the DCU calculates the corresponding parameter information for each cluster at each grid point. The parameter information includes: diffusion coefficient and reaction rate coefficient.

[0024] Thread synchronization ensures that all threads reach the same point in time and execute the same steps.

[0025] Calculate the initial reaction rate of each cluster at each grid point in the next time step;

[0026] Calculate the partial derivatives of each cluster at each grid point in the next time step for solving the equations;

[0027] Thread synchronization ensures that all threads reach the same point in time and execute the same steps.

[0028] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0029] 1) Expanding the spatiotemporal scale of cluster dynamics simulation methods

[0030] This embodiment implements a space-dependent cluster dynamics model and a DCU-CPU heterogeneous hybrid architecture version, which can perform large-scale parallel simulations on both CPU and DCU, expanding the spatiotemporal scale of cluster dynamics simulation methods.

[0031] 2) Reduce the computation time for cluster dynamics simulations

[0032] Leveraging the computing power of DCU, a parallel computing method for cluster dynamics on a DCU-CPU heterogeneous hybrid architecture was designed, effectively improving the execution efficiency of the method and significantly shortening the solution time of cluster dynamics. Compared with the known programs for cluster dynamics simulation under the same model, the performance is improved by 60%, thus solving the problem of excessively long computation time of existing space-dependent cluster dynamics simulation methods, and bringing the computation time to an acceptable range for practical applications. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating the parallel computing method for a cluster dynamics method for a DCU-CPU hybrid architecture provided in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the data structure on the DCU provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of DCU computing task partitioning provided in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the DCU-side computing process provided in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, this embodiment of the invention provides a parallel computing method for cluster dynamics of a DCU-CPU hybrid architecture, including:

[0040] S1, establish the data structure information on the DCU to store the data required for computation;

[0041] like Figure 2 As shown, the DCU (e.g., the Dawning DCU) mainly contains the following four data structures: a data structure for basic cluster information, a management structure for cluster information, a management structure for grid information, and a data structure for basic cluster response information. Figure 2 The dashed lines with arrows indicate the data structures pointed to by the stored address information. To the right of the four structures is a portion of the data information contained within them, specifically:

[0042] 1. Data structure of basic cluster information, including but not limited to one or more of the following information for a single type of cluster: size, number, diffusion coefficient, binding energy, migration energy, reaction radius, and concentration.

[0043] 2. The cluster information management structure includes, but is not limited to, one or more of the following: lattice size of the simulated object, defect atom size, reaction information storage address, number of reactions, and cluster information storage address. The cluster information storage address points to the structure described in 1, and the reaction information storage address points to the structure described in 4.

[0044] 3. The management structure for grid information includes, but is not limited to, one or more of the following: number of grids, grid size, and grid temperature.

[0045] 4. A data structure for basic information about cluster reactions, including but not limited to one or more of the following: reaction type, storage address of cluster information participating in the reaction, and reaction rate coefficient, wherein the storage address of cluster information participating in the reaction points to the structure described in 1.

[0046] S2, the CPU transmits the data required for computation to the DCU;

[0047] In this embodiment, the CPU uses HIP programming language to copy the initialized cluster basic information, cluster information management structure information, grid information management structure information, and cluster reaction basic information from the CPU memory to the corresponding structures in the DCU memory.

[0048] S3, assign the task of solving the established grid region to the DCU terminal;

[0049] In this embodiment, the solution domain is decomposed into a grid. During the calculation process, each grid needs to calculate information such as the concentration of each cluster on the grid and the rate coefficient of the reaction. In this embodiment, the calculation task on a single grid is taken as the calculation unit, which is the smallest task division unit and can correspond to the smallest calculation unit (computation thread) on the DCU. Therefore, the calculation tasks on a single grid are allocated to threads on the DCU, such as... Figure 3As shown, a task partitioning method is formed, consisting of a grid-thread and a region-thread pool; where, Figure 3 This is a schematic diagram of DCU computing task partitioning. From top to bottom, it shows the overall idea of ​​computing task partitioning, which corresponds the simulation area to the DCU structure. First, the area after meshing is further divided, with one area corresponding to one DCU computing structure (CU). Then, each grid point in each area corresponds to a computing thread in the CU.

[0050] S4, on the CPU side, calls the solver to establish the initial values ​​and equation matrix for solving the system of equations, and then solves the system of equations.

[0051] In this embodiment, the use of the PETSc (Portable Extensible Toolkit for Scientific Computation) external solver as an example is taken to illustrate this part. However, it should be noted that other external solvers can also be used to execute the corresponding process in this embodiment. The TS time-stepping integral solver in PETSc is used to perform time-dependent cluster dynamics simulation. The Vec and Mat structures in PETSc are used to establish the initial values ​​and equation matrix of the equation system.

[0052] S5 transmits the solution results of the system of equations to the DCU.

[0053] In this embodiment, the solution results of the equation system, including the information on the concentration of each defect at each grid point and the temperature, are transmitted to the DCU using HIP programming statements to update the data of the relevant structures on the DCU.

[0054] S6, the DCU calculates the initial values ​​(the initial reaction rate described in A3) and the equation matrix (the partial derivatives described in A4) required for the next time step based on the received solution results and the data needed for the calculation. Figure 4 As shown, the specific steps may include:

[0055] A1, the DCU terminal calculates the corresponding parameter information of each cluster at each grid point based on the received solution results. The parameter information includes: diffusion coefficient and reaction rate coefficient.

[0056] A2, thread synchronization, ensures that all threads reach the same point in time and execute the same steps;

[0057] A3, calculate the initial value of the reaction rate of each cluster at each grid point in the next time step;

[0058] A4, calculate the partial derivatives of each cluster at each grid point in the next time step for solving the equation;

[0059] A5, thread synchronization, ensures that all threads reach the same point in time and execute the same steps.

[0060] In this embodiment, steps A1-A5 can be implemented by using the HIP programming language to start the DCU core function calculation part.

[0061] S7, the DCU sends the calculation results (including the initial values ​​and equation matrix calculated in S6) back to the CPU.

[0062] Figure 4 This is a flowchart of the DCU-side computation process, which is roughly divided into three parts. The first part transmits the solution results to the DCU, where the DCU updates the data and performs thread synchronization. The second part calculates the initial values ​​and partial derivatives required for the next time step and performs thread synchronization. The third part transmits the computation results from the DCU to the CPU.

[0063] In this embodiment, the data required for the next time step is copied from the DCU memory to the corresponding structure in the CPU memory using the HIP programming language.

[0064] S8, repeat S4-S7 until the required computation time for the simulation is reached.

[0065] In this embodiment, the main flow of the program is the loop section S4-S7. The program uses whether the simulation time has reached the predetermined simulation time requirement as the criterion for judgment and performs iterative loops. In each step of the loop, the simulation time is increased by a time step. The detailed process of the DCU-side calculation in the loop is as follows: Figure 4 As shown.

[0066] In summary, the parallel computing method for cluster dynamics of DCU-CPU hybrid architecture provided by the embodiments of the present invention has at least the following beneficial effects:

[0067] 1) Expanding the spatiotemporal scale of cluster dynamics simulation methods

[0068] This embodiment implements a space-dependent cluster dynamics model (including data structure design and architecture-independent parts of the computation process, such as the calculation of reaction rate coefficients, although performed on the DCU in this invention, this step is itself part of the cluster dynamics model, and performing it on the DCU is an optimization of the computation task), and implements a DCU-CPU heterogeneous hybrid architecture version, which can perform large-scale parallel simulations on both CPU and DCU, expanding the spatiotemporal scale of the cluster dynamics simulation method.

[0069] 2) Reduce the computation time for cluster dynamics simulations

[0070] Leveraging the computing power of DCU, a parallel computing method for cluster dynamics on a DCU-CPU heterogeneous hybrid architecture was designed, effectively improving the execution efficiency of the method and significantly shortening the solution time of cluster dynamics. Compared with the known programs for cluster dynamics simulation under the same model, the performance is improved by 60%, thus solving the problem of excessively long computation time of existing space-dependent cluster dynamics simulation methods, and bringing the computation time to an acceptable range for practical applications.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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.

Claims

1. A parallel computing method for cluster dynamics in a DCU-CPU hybrid architecture, characterized in that, include: S1, establish data structure information on the DCU to store the data required for computation; wherein, the data structure includes: a data structure for basic cluster information, a management structure for cluster information, a management structure for grid information, and a data structure for basic cluster reaction information; S2, the CPU transmits the data required for computation to the DCU; S3, assign the task of solving the established grid region to the DCU terminal; S4, on the CPU side, calls the solver to establish the initial values ​​and equation matrix for solving the system of equations, and then solves the system of equations. S5 transmits the solution results of the system of equations to the DCU. S6, the DCU calculates the initial values ​​and equation matrix required for the next time step based on the received solution results and the data required for the calculation. S7, the DCU end transmits the calculation results back to the CPU end; S8, repeat S4-S7 until the required simulation time is reached; The data structure of the basic information of the clusters includes one or more of the following: size, number, diffusion coefficient, binding energy, migration energy, reaction radius and concentration of a single type of cluster; The cluster information management structure includes one or more of the following: lattice size of the simulated object, defect atom size, reaction information storage address, number of reactions, and cluster information storage address; wherein the cluster information storage address points to a data structure of basic cluster information, and the reaction information storage address points to a data structure of basic cluster reaction information. The management structure for the grid information includes one or more of the following: number of grids, grid size, and grid temperature; The data structure for the basic information of the cluster reaction includes one or more of the following: reaction type, storage address of cluster information participating in the reaction, and reaction rate coefficient; wherein the storage address of cluster information participating in the reaction points to the data structure for the basic information of the cluster. The DCU terminal calculates the initial values ​​and equation matrix required for the next time step based on the received solution results and the received data required for the calculation, including: Based on the received solution results, the DCU calculates the corresponding parameter information for each cluster at each grid point. The parameter information includes: diffusion coefficient and reaction rate coefficient. Thread synchronization ensures that all threads reach the same point in time and execute the same steps. Calculate the initial reaction rate of each cluster at each grid point in the next time step; Calculate the partial derivatives of each cluster at each grid point in the next time step for solving the equations; Thread synchronization ensures that all threads reach the same point in time and execute the same steps.

2. The parallel computing method for the cluster dynamics method for DCU-CPU hybrid architecture according to claim 1, characterized in that, The process of assigning the established mesh region solution task to the DCU includes: The solution domain is decomposed by meshing, and the computational tasks on a single mesh are used as computational units and allocated to threads on the DCU, forming a task partitioning method of mesh-thread and region-thread pool.

Images