Access the full text.

Sign up today, get DeepDyve free for 14 days.

Applied Sciences
, Volume 12 (20) – Oct 11, 2022

/lp/multidisciplinary-digital-publishing-institute/accelerated-parallel-numerical-simulation-of-large-scale-nuclear-IKaG1X1l1s

- Publisher
- Multidisciplinary Digital Publishing Institute
- Copyright
- © 1996-2022 MDPI (Basel, Switzerland) unless otherwise stated Disclaimer The statements, opinions and data contained in the journals are solely those of the individual authors and contributors and not of the publisher and the editor(s). MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Terms and Conditions Privacy Policy
- ISSN
- 2076-3417
- DOI
- 10.3390/app122010193
- Publisher site
- See Article on Publisher Site

applied sciences Article Accelerated Parallel Numerical Simulation of Large-Scale Nuclear Reactor Thermal Hydraulic Models by Renumbering Methods † † Huajian Zhang , Xiao-Wei Guo , Chao Li , Qiao Liu, Hanwen Xu and Jie Liu * Institute for Quantum Information & State Key Laboratory of High Performance Computing, College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, China * Correspondence: liujie@nudt.edu.cn † These authors contributed equally to this work. Abstract: Numerical simulation of thermal hydraulics of nuclear reactors is widely concerned, but large-scale ﬂuid simulation is still prohibited due to the complexity of components and huge computational effort. Some applications of open source CFD programs still have a large gap in terms of comprehensiveness of physical models, computational accuracy and computational efﬁciency compared with commercial CFD programs. Therefore, it is necessary to improve the computational performance of in-house CFD software (YHACT, the parallel analysis code of thermohydraulices) to obtain the processing capability of large-scale mesh data and better parallel efﬁciency. In this paper, we will form a uniﬁed framework of meshing and mesh renumbering for solving ﬂuid dynamics problems with unstructured meshes. Meanwhile, the effective Greedy, RCM (reverse Cuthill-Mckee), and CQ (cell quotient) grid renumbering algorithms are integrated into YHACT software. An important judgment metric, named median point average distance (MDMP), is applied as the discriminant of sparse matrix quality to select the renumbering methods with better effect Citation: Zhang, H.; Guo, X.-W.; Li, for different physical models. Finally, a parallel test of the turbulence model with 39.5 million C.; Liu, Q.; Xu, H.; Liu, J. Accelerated grid volumes is performed using a pressurized water reactor engineering case component with Parallel Numerical Simulation of Large-Scale Nuclear Reactor Thermal 3*3 rod bundles. The computational results before and after renumbering are also compared to Hydraulic Models by Renumbering verify the robustness of the program. Experiments show that the CFD framework integrated in Methods. Appl. Sci. 2022, 12, 10193. this paper can correctly perform simulations of the thermal engineering hydraulics of large nuclear https://doi.org/10.3390/ reactors. The parallel size of the program reaches a maximum of 3072 processes. The renumbering app122010193 acceleration effect reaches its maximum at a parallel scale of 1536 processes, 56.72%. It provides a Academic Editors: Vasily Novozhilov basis for our future implementation of open-source CFD software that supports efﬁcient large-scale and Cunlu Zhao parallel simulations. Received: 19 September 2022 Keywords: mesh renumbering; thermal engineering hydraulics; CFD; YH-ACT; RCM; parallel Accepted: 4 October 2022 calculation Published: 11 October 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional afﬁl- 1. Introduction iations. Because of the importance of nuclear reactions and the complexity of their physical models, research exploration using large-scale numerical simulations is essential. Some early studies of nuclear reactor thermodynamics were based on several parameters at large scale or one-dimensional system analysis methods, which mainly studied the average Copyright: © 2022 by the authors. characteristics of physical quantities in nuclear power systems. With the boom in high Licensee MDPI, Basel, Switzerland. performance computers, nuclear power thermohydraulic studies based on more reﬁned This article is an open access article computational ﬂuid dynamics (CFD) methods and thermohydraulic software have been distributed under the terms and widely studied and applied [1,2]. conditions of the Creative Commons The internal structure of nuclear reactors is complex, and the dominant reactor type Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ today is the pressurized water reactor. Core fuel rod bundles are an important component 4.0/). of a pressurized water reactor. The fuel rods are surrounded by a ﬂowing liquid coolant. Appl. Sci. 2022, 12, 10193. https://doi.org/10.3390/app122010193 https://www.mdpi.com/journal/applsci Appl. Sci. 2022, 12, 10193 2 of 16 There is a violent ﬂow of liquid between the bar bundles and between the bar bundles and the walls, and they are responsible for the transfer of momentum and heat. It is widely believed that the ﬂow pulsations in the gap are closely related to the large-scale vortices. It has been studied that vortexes will form in pairs on both sides of the slit [3,4]. The ﬂow and heat transfer characteristics of coolant ﬂowing through the core of a nuclear reactor are extremely complex. To analyze the turbulent motion and thermodynamics of large-scale fuel rod bundles [5–10], large-scale CFD numerical simulations based on FVM (ﬁnite volume method) are a widely accepted solution. However, large-scale CFD numerical simulations require an extremely large amount of mesh data. The usual solution is to encrypt the mesh of the physical model. The larger the amount of data, the larger the solution size and the accuracy of the solution are naturally improved. However, this will lead to more and more complex and difﬁcult to solve calculations. Today, the use of large-scale CFD techniques in thermal hydraulic analysis is still very difﬁcult due to high technical requirements and shortage of computational resources [11]. In the CFD, simulation of ﬁne ﬂow ﬁeld structures in real-world engineering cases is still forbidden [12]. As a result, there is widespread interest in improving computational efﬁciency by well-developed parallel technology. The CFD solving process involves three parts: pre-processing, numerical solving, and post-processing. The use of appropriate meshing methods and mesh numbering in the preprocessing process will greatly improve the computational performance of ﬂuid dynamics simulations. Taking FVM [13,14] as an example, after meshing the physical model, assembly of linear systems in parallel from the number of grid cells of each divided component is necessary. Then, solving them iteratively,. In addition, the numbering of the grid determines the order of access to the individual processes for parallel computation and affects the cache hit rate. Since modern computer architecture is characterized by multiple storage levels, reasonable grid numbering has a signiﬁcant impact on the efﬁcient utilization of computational performance. One of the key technologies in CFD preprocessing is the meshing technique. Its theoretical basis is becoming more and more mature. In recent years, the research ﬁeld of ﬁnite element meshing has shifted from two-dimensional plane problems to three- dimensional entities, and the research focus has changed from triangular (tetrahedral) meshes to quadrilateral (hexahedral) meshes, focusing on the fully automatic generation of meshes, mesh adaption and other research. Mesh partitioning can be abstracted as graph partitioning. The relationship of cells and neighbors can be abstracted as the relationship of vertices and edges. The large-scale physical model is divided into multiple blocks by coarse- grained and ﬁne-grained lattice partitioning, which facilitates parallel computation [15]. It is a challenge to achieve high quality meshing as quickly as possible and under the condition that the load balance is satisﬁed. The efﬁciency and cell quality of meshing need to be further improved. Another key technique in preprocessing is the grid numbering technique. The effect of grid numbering techniques on the computational efﬁciency of solving sparse linear systems is a classical topic with many research results, examples of which can be found in the work of Gibbs et al. [16], Cuthill and McKee [17], Sloan [18], Akhras and Dhatt [19], and George et al. are reﬂected in the work of Reverse Cuthill–Mckee [20]. Because the heuristic renumbering algorithm consumes a lot of time during preprocessing. Therefore, heuristic algorithms may be more appropriate when computational performance is not a critical issue [21,22]. In addition, this paper studies how to accelerate the computational performance of general-purpose CFD software, so heuristic algorithms are not considered for the time being. In general, meshing and grid numbering techniques are a traditional approach to performance optimization, and there are many existing methods to choose from. However, these techniques still lack quantitative analysis for large-scale thermal-hydraulic cases. In particular, the optimization effectiveness for general-purpose CFD software executed on supercomputers is still lacking. Therefore, this paper focuses on the effectiveness of the FVM-based computational framework for use in unstructured grids, integrating Appl. Sci. 2022, 12, 10193 3 of 16 typical meshing methods and mesh numbering techniques. Ultimately, a suitable mesh renumbering method is selected to enhance the parallel scale and computational speed of numerical simulation. The main innovative points of this paper are as follows: (1) A general framework for further integration of parallel meshing methods and renumbering schemes in general- purpose CFD software for unstructured meshes; and (2) improved runtime speed and parallel scale for real-world numerical simulations of large-scale complex cases. The case is a thermal hydraulic analysis of a 3*3 fuel rod bundle with 39.5 million grid volumes. The structure of this paper is as follows: Section 2 describes the generic CFD soft- ware used in this paper and aims to demonstrate the advantages of the software between data structure and parallel communication. Section 3 shows the meshing strategy and the integrated RCM renumbering algorithm. Section 4 demonstrates the correctness of the massively parallel numerical simulations before and after renumbering and the huge improvement of the renumbering algorithm on the numerical simulation efﬁciency. Con- clusions are included in Section 5. 2. General CFD Software Based on Finite Volume Method The general-purpose CFD software used in this paper, YHACT, was initially used for thermal-hydraulic analysis. YHACT uses a modular development architecture based on scalability. It is based on a framework blueprint of CFD program modules consisting of key steps in CFD solving such as data loading, physical pre-processing, iterative solving, and result output. YHACT mainly addresses the thermal-hydraulic CFD program requirements for nuclear reactors, and more details about it can be found in [23]. 2.1. Scalability of YHACT Software Architecture Thanks to the modern features of object-oriented technology, YHACT can be developed modularly and has ﬂexible scalability. YHACT is based on common parallel algorithms and parallel support, combined with existing advanced software engineering methods. It uses physical model module, numerical discretization module, linear equation system module, and parallel communication module as the main components to form the general framework of CFD program based on supercomputing system [24]. The pre-processing toolbox has functions such as mesh control and physical pre- processing. The physical model and numerical solution constitute the CFD solver, encap- sulating the physical mathematical model and efﬁcient mathematical solution methods with parametric conﬁguration capabilities. The common algorithm implements parallel computation such as sparse iterative methods and preconditioners. Through close cou- pling with the computer architecture, YHACT can obtain support for massively parallel computational optimization and provide fundamental computational power for the solver. Parallel support can encapsulate the structural features of high-performance computer systems and provide fundamental computing, communication, storage, and debugging functions, thereby supporting the efﬁcient operation of large-scale CFD computing appli- cations. Therefore, the good scalability of YHACT makes it easier to integrate meshing techniques and mesh rearrangement schemes. As shown in Figure 1, it does not require much change to the software architecture, but only requires the implementation of the required functions by calling the interfaces of each module. The overall framework of the grid partitioning technique and grid rearrangement scheme will be presented in Section 3. 2.2. Parallel Solving Algorithm 2.2.1. Parallel Decomposition of Grid Data A simple and intuitive way of parallelism is to decompose the grid data. Similar to Cartesian partitioning, the grid data are divided into as many copies as required and then distributed to different processes for computation. The processes rely on communication to exchange data between them. Appl. Sci. 2022, 12, 10193 4 of 16 Figure 1. Overall framework of YHACT, framework after integration of meshing and mesh renum- bering functions. Red boxes indicate new modules for integration. The data decomposition is shown in Figure 2. The grid to be solved is divided into non-overlapping blocks of grid sub-cells, and each process reads only one piece of grid data. After data decomposition, dummy cells are generated on the physical boundaries adjacent to each sub-grid and other grids for data communication only. Dummy cells are only used for data exchange and are not part of the physical model, which makes each cell consider only the data of its neighbors. The data on the dummy cell can be used as boundary data. This simple data decomposition does not share the grid data. In addition, a discrete format of up to 2nd order is used, and only the values of adjacent grid cells are used in the grid calculation. Therefore, the cross-process data to be acquired can be stored on the boundary line (dummy cell in Figure 2). The solution can be divided into the following three steps: (1) boundary initialization (fetching data on adjacent processors); (2) process local computation; and (3) boundary update (updating data on adjacent processors). Appl. Sci. 2022, 12, 10193 5 of 16 Figure 2. Grid data decomposition. The left ﬁgure shows the original grid model. The right side indicates the four sub-grids after data decomposition. Dummy cells are generated at the boundary of each subgrid. It is only used for data communication during grid computation of different processes. 2.2.2. Communication Patterns after Parallel Decomposition Since each grid block only needs to communicate with its neighbors. The ﬁrst step in establishing the communication relationship is to ﬁnd all the neighboring grid blocks and store them in the local array. Each grid block then only has information about itself and its neighbors, which are “localized” grid blocks. When the parallel scale increases, the current grid only needs to traverse the “localized” grid blocks when communicating, which greatly improves the efﬁciency of communication. MPI asynchronous communication is also used to further improve communication efﬁciency. An example of a communication model is shown in Figure 3. Figure 3. Example of communication between four processes. The top left indicates the subgrid to be calculated. The upper right indicates the three different process states during communication. The bottom indicates the state of each process at the same time step. Note that each process performs a time step of computation after receiving data, hence skip. Appl. Sci. 2022, 12, 10193 6 of 16 2.2.3. Parallel Architecture Based on MPI and OpenMP Based on the grid model decomposition described above, two or more levels of data partitioning can be carried out. Depending on the computational resources or data structure, a certain number of “sub-grids” can be used as input to a process to perform computation. A hybrid MPI/OpenMP parallelism mode can be used. MPI parallelism is used on the ﬁrst-level partition, and then the ﬁrst-level partition is divided into several second-level sub-partitions. The number of second-level sub-partitions is generally equivalent to the cores inside the compute node. Coarse-grained OpenMP parallelism is performed on the second-level partition. This multi-tier parallel architecture is adapted to the current architecture of supercom- puter clusters, where primary partitions are assigned to individual compute nodes, while secondary partitions are on top of the same compute node. The above is also applicable to heterogeneous systems. The ﬁrst layer still uses MPI communication based on the grid’s ﬁrst-level partition interfacing information.The second layer is OpenMP parallelism, i.e., coarse-grained parallelism between compute cores, based on secondary subpartitions of the grid. Data transfer between CPUs and coprocessors is performed using the corresponding channels. When there is no coprocessor, the above heterogeneous parallel mode auto- matically degenerates to homogeneous parallel mode. In order to provide better control over the different partitioning layers, the whole computational domain is divided into two layers: Layer 1 is the Region partition and Layer 2 is the Zone sub-partition. In parallel computation, the Region partition is assigned to a compute node, corresponding to an MPI process, and the Zone subpartition is assigned to a compute core thread, corresponding to an OpenMP thread. A typical CFD solver execution process mainly includes residual initialization, time step calculation, ﬂux calculation, time advancement, ﬂow ﬁeld update, and intersection communication. The solver is applied to the computational domain Zone. For different physical models, multiple solvers may be loaded on a Zone. During the computation, Region traverses each of these Zones in turn, calling the corresponding solver to perform the computational task. When OpenMP parallelism is used, the computational tasks can be performed concurrently for each Zone. 2.2.4. Parallel Communication for Data Packaging In the multi-level parallel architecture described above, there may be multiple subgrid blocks on each MPI process. If we directly follow the above model of cyclic traversal of grid blocks for MPI/OpenMP hybrid parallel communication, there will be the problem of multiple communications between processes. Take Figure 4 as an example, Region0 and Region1 have 2 sets of adjacent subgrid blocks, and each exchange of data requires two send and receive processes. Therefore, the grid partition data between processes can be packaged as shown in Figure 5 to reduce the number of communications. Figure 4. Partition level of two processes. Appl. Sci. 2022, 12, 10193 7 of 16 (a) (b) Figure 5. Example of packaging of process data. (a) unpacked transfer of data; (b) packet transmission of data. 3. A General Framework for Grid Partitioning Techniques and Grid Renumbering Methods 3.1. Parmetis-Based Grid Partitioning Strategy Metis and ParMetis are tools that work on physical model partitioning. Usually, using the Metis library for mesh slicing of physical models using a serial strategy is an ideal solution. However, it is slow under large-scale mesh partitioning. ParMetis is a parallel version of Metis, an MPI parallel library. It has many built-in algorithms for mesh partitioning and repartioning of unstructured graphs. ParMetis started out as a tool for large-scale graph partitioning, but mesh models in CFD can also be abstracted to graphs. Therefore, ParMetis is particularly well suited to deal with unstructured meshing problems in large scale numerical simulations. Simulations of ﬂuid problems often require thousands of cluster nodes to be computed simultane- ously. Before the computation, a regional decomposition of the computational grid has to be performed.Communication is required between different regions in the computation. Therefore, grid region decomposition has a greater impact on load balancing and inter- process communication. PartMetis enables the partitioned area grid to be approximately the same and the number of intersections to be minimized. Therefore, the amount of data exchanged is minimized, which in turn signiﬁcantly reduces the communication time between processes. ParMetis provides the ParMETIS_V3_PartGeomKway routine for computing par- titions of graphs derived from ﬁnite element meshes where vertices have coordinates associated with them. Given a graph distributed between processor and vertex coor- dinates, ParMETIS_V3_PartGeomKway uses the space-ﬁlling curve method to quickly compute an initial partition. The graph is redistributed according to this partition and ParMETIS_V3_PartKway is then called to compute the ﬁnal high quality partition. Infor- mation about the partitioning process of Parmetis is available in [25]. 3.2. Grid Renumbering Framework The neighbor relationship between adjacent cells reﬂects the memory access order of the numerical simulation. The system of discrete equations is obtained by discretizing on each interval of the control volume of the entire domain to be solved. Discrete systems of equations are presented as sparse matrices in a computer [13], where the non-zero elements of each row denote the coefﬁcients of the unknown quantities of the current set of equations. The matrix usually contains a few nonzero elements , and the number of non-zero elements is exactly equal to the number of elements associated with this row. For different grid types, the coefﬁcient matrix of a structured grid is banded, while the coefﬁcient matrix of an unstructured grid is relatively irregular and only has the characteristics of a diagonal matrix. Therefore, based on the above characteristics, many solutions have been proposed. The methods for solving large-scale equations can be divided into direct and iteration approaches. Since the real case grid size of thermal hydraulics is usually large, the solution process of direct method requires a large amount of computational resources and the cost of solving the inverse is very high. Therefore, Appl. Sci. 2022, 12, 10193 8 of 16 iteration approaches are usually used to addressing massive thermal-hydraulic problems. The iterative method means that the algorithm needs to be used several times to reach a convergence state. For example, ILU decomposition method, and the conjugate gradient method can be used as representatives of the iterative method for solving. The key to improving the speed of solving is that, by writing part of the memory data to the cache, the solver can access the required solved data with higher efﬁciency. In other words, by rearranging the mesh without altering the program structure [26], more valid data can be written to the cache to prevent frequent memory reads for a higher cache hit rate. Therefore, this section integrates part of the grid renumbering framework in YHACT, including but not limited to the Greedy algorithm and CQ algorithm implementation, which can be found in [27], and the RCM algorithm is elaborated below. 3.2.1. RCM Algorithm After careful algorithm design, the RCM algorithm was shown to be able to adapt to large-scale distributed memory parallelism [28] and has been widely used in software such as OpenFoam and MatLab. When the numerical stage of sparse matrix computation is reached, the matrix is usually distributed. At this point, the RCM algorithm can improve the sparse structure of the sparse matrix, making the non-zero elements more concentrated, which can theoretically improve the storage efﬁciency of the computer. The RCM algorithm uses BFS to traverse the nodes in the graph. At each layer of the graph, nodes are sorted according to their degree labels for the renumbering of that layer. In this paper, due to the modular nature of the numerical simulation framework and the complexity of the case, the algorithm’s feature of sorting nodes according to their degrees is weakened to reduce the time cost of partially traversing the nodes. Algorithm 1 shows the overall program structure. Algorithm 1 RCM numbering algorithm input: mesh output: mesh 1: Select the initial unit to be numbered, set to mesh . Set its number to the value of the maximum number of units 2: Arrange the unnumbered neighbors of mesh in the container L according to the de- scending order of degree 3: Set the ﬁrst element of L to mesh and renumber it. The value is the maximum number of cells minus one 4: Implementation Step 2 5: Implementation of step 3, and the numbered values in decreasing order 6: Repeat until the numbering is complete Note that the initial numbered cells of the original RCM algorithm are selected by traversing the entire grid of the computational solution domain. The degree of meshes is used to achieve the desired optimization. Here, we require that the initial elements be selected as boundary cells or close to the boundary. This method prevents excessive time overhead due to the large and complex size of the case. It makes the algorithm more adaptive in increasingly complex cases. The key step is the numbering. The RCM algorithm sorts the meshes in reverse according to the sequence of degrees, from “nElement -1" to “0". 3.2.2. Selection of Grid Renumbering Methods Usually the sparse structure of the generated sparse matrices will vary for different renumbering methods. Bandwidth and proﬁle are two common methods to discriminate the quality of sparse matrices. In the computer ’s storage structure, the sparse matrix eventually participates in the CFD solver computation, and its sparse structure has a signiﬁcant impact on the cache miss rate during the computation, which in turn affects the computational efﬁciency of the numerical simulation. Nevertheless, the above only represents the sparse Appl. Sci. 2022, 12, 10193 9 of 16 structure of the matrix as a whole and ignores the degree of aggregation around the diagonal elements of the sparse matrix. The more elements close to the diagonal, the smaller the cache miss rate will be. Therefore, this paper adopts the MDMP discriminant method, and more details can be found in [27]. This ensures that more data are hit in the same cache: The MDMP discriminant formula is: d = (1) k=1 It has been shown that the MDMP metric can describe the dispersion of this sparse matrix very well. 4. Parallel Experiments for Numerical Simulation of Large Scale Nuclear Reactor Thermodynamics This section runs a typical conﬁguration on a high performance machine in modern times. The computer consists of about 120 computing nodes. Details of the HPC are shown in Table 1. The hardware conﬁguration CPU is an Intel(R) Xeon(R) Gold 6240R processor with 48 cores and 192G RAM. Table 1. The platform conﬁguration used in the study. Parameter Conﬁguration Processor model Intel(R) Xeon(R) Gold 6240R @2.40GHz Number of cores 48 cores Memory 192G OS Red Hat 8.3.1-5 Compiler GNU Compiler Collection (GCC 8.3.1) 4.1. CFD Physical Modeling of Nuclear Reactor Thermal Hydrology In the process of nuclear reactor thermodynamic analysis, the ﬂow of coolant heat transfer process needs to be modeled. Figure 6 illustrates the thermal hydraulic model of a nuclear reactor that is the main focus of this paper. The theoretical model needed to completely describe the ﬂow in the ﬂuid as well as the heat transfer process contains three types of conservation equations for mass, momentum and energy [29]. Figure 6. Fluid computing domain. Appl. Sci. 2022, 12, 10193 10 of 16 The structure inside the pressure vessel is complex, the range of ﬂow scales is large, and the practical application of engineering is more concerned with the average amount of physical quantities. Therefore, Reynolds averaging is generally performed for the three major equations to calculate the average physical quantities of ﬂuid ﬂow. In addition, then the turbulence model needs to be added in order to make the averaged mass, momentum and energy equation set closed. This is used to perform simulations of core cooling ﬂuid ﬂow and heat transmission . Therefore, for the thermodynamic model, a smaller grid size will lead to insufﬁcient accuracy of its simulation. Therefore, in this paper, a 3*3 assembly with about 39.5 million grid size is simulated numerically for a fuel rod bundle. The SIMPLE iterative algorithm and PETSc solver are used to perform the calculations. The relevant parameters of the calculation, such as the relaxation factor, are adjusted. Pressure relaxation factor and velocity relaxation factor are set to 0.3 and 0.5, respectively, and a turbulence physical quantity relaxation factor of 0.5 are used to calculate the turbulent ﬂow between fuel rod bundles with stirred wings. It is worth noting that YHACT currently mainly considers the implementation of the Reynolds time-averaged turbulence model. As for the ﬂuid–solid coupled heat transfer model, the heat transfer model between ﬂuid–solid domains is generally established through the intersection of ﬂuid and solid domains. When calculating the energy equation in the solid domain, the relevant physical quantities at the intersection of the ﬂuid domain are used as boundary conditions. In addition, when calculating the energy equation in the ﬂuid domain, the heat (temperature) distribution at the solid intersection is used as the source term of the ﬂuid energy equation to ﬁnally realize the coupled ﬂuid-solid heat transfer. Since this paper mainly concentrates on the turbulent heat transfer model, the solid models of fuel rod and control rod are neglected. The boundary condition of the solid interface on the pure ﬂuid is a ﬁxed heat ﬂux. This heat ﬂux is roughly calculated based on the solid bar size and the volumetric heat source at the time of ﬂuid–solid coupling. 4.2. Select the Grid Renumbering Algorithm This subsection focuses on how to select the appropriate grid renumbering algo- rithm. Based on the integrated renumbering strategy, the sparse structure of the sparse matrices generated by each type of grid renumbering is discriminated by MDMP metrics for the above fuel rod bundle model. Finally, the most suitable algorithm is selected for numerical simulation. Due to the large grid size, it is more difﬁcult to obtain the grid numbering geometry information of the physical model by serial means. Therefore, this subsection collects the numbering geometry information of the subgrids in process 0 for the two parallel scales of 48 and 96 processes to determine. Table 2 shows the sparse matrices generated by four different numbering methods for each of the two parallel scales. Below, each row of the sparse matrix corresponds to its MDMP index. A simple comparison can be made to conclude that the original numbering of the grid drawn by the ICEM library is confusing. After renumbering, the properties of the sparse matrix are better improved. Meanwhile, the RCM algorithm will be better compared to several other numbering methods. Therefore, in the following parallel tests, the RCM renumbering algorithm is chosen. Appl. Sci. 2022, 12, 10193 11 of 16 Table 2. The sparse matrix of the grid after being processed by different renumbering algorithms. The blue dots indicate the non-zero elements in the matrix. Numbering Strategy 48 Process 96 Process original MDMP 32,387.88 16,588.94 RCM MDMP 44.61 44.35 greedy MDMP 55.24 54.16 CQ MDMP 838.29 378.98 4.3. Massively Parallel Performance Testing 4.3.1. Proof of Correctness Parallel performance tests must ensure the correctness of the numerical simulations. After optimization by renumbering the modules, proof of program robustness is essential. Figure 7 illustrates the two slice models tested in this subsection. (a) shows the slice with Z = 0.01, (b) shows the slice with Z = 0.01. In addition, (b) belongs to the shelf section, which has more complex physical properties. Tables 3 and 4 show the proﬁle cloud plots for the slice at Z = 0.01 and the proﬁle cloud plots for the slice at Z = 0.01, respectively, for comparing the correctness of the numerical simulations before and after renumbering. Each table shows the contour plots of temperature, pressure, and velocity. The ﬁrst column represents the different physical quantities. The second column represents the range of physical quantities in the cloud plot. The third and fourth columns show the sliced cloud plots at the same locations before and after the renumbering algorithm. The white numbers in each cloud represent the values of physical quantities at the current position. When the same physical quantities are compared, both numbering methods take the same position. Therefore, the corresponding Appl. Sci. 2022, 12, 10193 12 of 16 positions of the white numbers before and after renumbering are the same. The white numbers in the cloud plot are the values of physical quantities at the same locations, which are used to determine whether the accuracy of the renumbering optimization meets the requirements and whether the robustness of the program is guaranteed. It can be seen that the renumbering does not change the correctness of the numerical simulation either at the fuel bar (Z = 0.01) or at the shelf (Z = 0.01). The robustness of the procedure is guaranteed. Table 3. Contour clouds for slices with Z = 0.01. The first column indicates the different physical quantities. The second column indicates the range of physical quantities in the cloud map. The third and fourth columns show the sliced clouds at the same locations before and after the renumbering algorithm. The white numbers in each cloud represent the values of physical quantities at the current position. When the same physical quantities are compared, both numbering methods take the same position. Physical Properties Range Old RCM Temperature Pressure Velocity Appl. Sci. 2022, 12, 10193 13 of 16 Table 4. Contour clouds for slices with Z = 0.01. The speciﬁc information is the same as the legend in Table 3. Physical Properties Range Old RCM Temperature Pressure Velocity (a) (b) Figure 7. Slicing model. (a) slice with Z = 0.01; (b) slice with Z = 0.01. In addition, slice (b) is located in the shelf. Appl. Sci. 2022, 12, 10193 14 of 16 4.3.2. Parallel Testing of Large-Scale Numerical Simulations Finally, in this paper, parallel numerical simulations are performed for a fuel rod bundle of 39.5 million grid size with a different number of processes. Table 5 shows the time to iterate 100 steps on different processes for the original grid and the grid renumbered by RCM. The ﬁrst column represents the amount of nodes used. The second column indicates the number of processes used, trying to ensure that each node is used 100%. The third and fourth columns indicate the total simulation time for the original mesh and the RCM renumbered mesh. The ﬁfth column indicates the speedup of the numerical simulation of the renumbered grid compared to the original grid. It can be obtained that the renumbering module has a good effect on the computational effectiveness of the mathematical simulation of the fuel rod technique. In addition, the parallel scale reaches a maximum of 3072 processes. At the parallel scale of 1536 processes, the renumbering module has the maximum speedup effect of 56.72%. In addition, at the parallel scale of 3072 processes, the time difference is not much. This may be due to the fact that, when the grid size is certain, the larger the parallel scale, the matrix computation time within the processes decreases dramatically, while the percentage of communication time increases dramatically. Table 5. The degree of enhancement of the RCM algorithm for numerical simulation of fuel rod bundles at different parallel scales. The ﬁrst column represents the number of nodes. The second column represents the amount of processes. The third and fourth columns indicate the total simulation time for the original grid and the grid renumbered by RCM. The ﬁfth column indicates the numerical simulation speedup of the renumbered grid compared to the original grid. Time No. of Nodes No. of Processes Speedup Rate Original(s) RCM(s) 2 96 19731.4 13682.5 30.66% 4 192 8635.3 6727.38 22.09% 8 384 3761.11 3342.41 11.13% 16 768 1794.29 1568.07 12.6% 32 1536 1618.58 700.528 56.72% 64 3072 748.43 716.175 4.31% 5. Conclusions In this paper, we focus on how to apply suitable renumbering algorithms to enhance the performance of large-scale nuclear reactor fuel rod technology in a CFD numerical sim- ulation framework. We do this by further integrating the meshing and mesh renumbering framework of YHACT, a general-purpose CFD software. The RCM algorithm is easily im- plemented by relying on the modular nature of the software. In this paper, we ﬁrst perform parallel meshing of the model using Parmetis to ensure that each process obtains a similar size of mesh blocks and thus try to ensure that the computation time of each process is not signiﬁcantly different. In addition, then, through rigorous experiments, we use MDMP as the discriminative index of sparse matrix quality to select a more suitable renumbering algorithm for the experimental case of this paper. Through the MDMP metric, it is easy to select the RCM algorithm is what we need. In addition, by plotting the contour clouds at the bar bundle and at the shelf, it is concluded in detail that our renumbering algorithm has a strong scalability in YHACT. This is also due to the modular development of YHACT, and the easy setup of the API. By observing the three contour plots of temperature, pressure, and velocity, it can be seen that the renumbered grid does not have any effect on the ﬁnal Appl. Sci. 2022, 12, 10193 15 of 16 solution results. The solution accuracy is well within the acceptable range. Therefore, we boldly and conﬁdently conducted a large-scale parallel test. In the end, the number of processes for the 3*3 fuel rod bundle problem reached 3072, and the most signiﬁcant speedup of 56.72% was achieved when the number of processes reached 1536. Meanwhile, the overall speedup increases to 56.72% when the speedup efﬁciency is gradually increased from 96 to 1536 threads. The acceleration efﬁciency does not decrease with the increase of parallelism size. Thus, the goodness of the strong scalability of parallelism was tested and proved. As industry develops and advances, the complexity of CFD numerical simulations will gradually increase. The gradual development of supercomputers will inevitably lead to an increasing parallel scale. It is necessary to improve the efﬁciency of numerical simu- lation through renumbering strategy. It is also necessary to select a suitable renumbering algorithm without affecting the scalability of the numerical simulation. It can be seen that the ﬁnal test results of this paper have good scalability and obvious acceleration effect.The difﬁculty lies in the easy combination of different acceleration methods. In addition, the best way to do this is to continuously enrich the functionality of the general CFD software through modular development. The signiﬁcance of this is that it does have a strong ac- celeration effect on real engineering cases. Continuously increasing the complexity and parallel scale of numerical simulations is the only way to better identify problems in the development process. Continuous problem solving leads to faster operation and more ac- curate numerical simulations. The next step is naturally to investigate which renumbering algorithm is suitable for different types of cases, such as turbulence, laminar ﬂow, steady state, transient, etc., in order to further improve the software solution accuracy, parallel scale, and running speed. Another direction is to dive into the Solver and explore how the renumbering algorithm affects the assembly and solution of the matrix. Author Contributions: Conceptualization, J.L. and X.-W.G.; methodology, H.Z.; software, X.-W.G., C.L., H.Z.; investigation, X.-W.G.; data curation, H.X. and Q.L.; writing—original draft preparation, H.Z.; writing—review and editing, H.Z.; funding acquisition, J.L., X.-W.G. and C.L. All authors have read and agreed to the published version of the manuscript. Funding: This work was partially supported by the National Natural Science Foundation of China (Grant Nos. 61902413, 12102468, and 12002380) and the National University of Defense Technology Foundation (No. ZK21-02, No. ZK20-52). Conﬂicts of Interest: no conﬂict of interest. References 1. Chen, R.; Tian, M.; Chen, S.; Tian, W.; Su, G.; Qiu, S. Three dimensional thermal hydraulic characteristic analysis of reactor core based on porous media method. Ann. Nucl. Energy 2017, 104, 178–190. 2. Wang, M.; Zuo, Q.; Yu, H.; Tian, W.; Su, G.; Qiu, S. Multiscale thermal hydraulic study under the inadvertent safety injection system operation scenario of typical pressurized water reactor. Sci. Technol. Nucl. Install. 2017, 2017, 2960412. 3. Krauss, T.; Meyer, L. Experimental investigation of turbulent transport of momentum and energy in a heated rod bundle. Nucl. Eng. Des. 1998, 180, 185–206. 4. Guellouz, M.; Tavoularis, S. The structure of turbulent ﬂow in a rectangular channel containing a cylindrical rod–Part 1: Reynolds-averaged measurements. Exp. Therm. Fluid Sci. 2000, 23, 59–73. 5. Kaiser, H.; Zeggel, W. Turbulent ﬂows in complex rod bundle geometries numerically predicted by the use of FEM and a basic turbulence model. In Proceedings of the Third International Topical Meeting on Reactor Thermal Hydraulics, Newport, RI, USA, 15–18 October 1985. 6. Lee, K.B.; Jang, H.C. A numerical prediction on the turbulent ﬂow in closely spaced bare rod arrays by a nonlinear k–" model. Nucl. Eng. Des. 1997, 172, 351–357. 7. Baglietto, E.; Ninokata, H. A turbulence model study for simulating ﬂow inside tight lattice rod bundles. Nucl. Eng. Des. 2005, 235, 773–784. 8. Baglietto, E.; Ninokata, H.; Misawa, T. CFD and DNS methodologies development for fuel bundle simulations. Nucl. Eng. Des. 2006, 236, 1503–1510. 9. Cheng, X.; Tak, N. CFD analysis of thermal–hydraulic behavior of heavy liquid metals in sub-channels. Nucl. Eng. Des. 2006, 236, 1874–1885. Appl. Sci. 2022, 12, 10193 16 of 16 10. Guo, R.; Oka, Y. CFD analysis of coolant channel geometries for a tightly packed fuel rods assembly of Super FBR. Nucl. Eng. Des. 2015, 288, 119–129. 11. Hu, R.; Fanning, T. Progress Report on Development of Intermediate Fidelity Full Assembly Analysis Methods; Technical report; Argonne National Lab. (ANL): Argonne, IL, USA, 2011. 12. Slotnick, J.P.; Khodadoust, A.; Alonso, J.; Darmofal, D.; Gropp, W.; Lurie, E.; Mavriplis, D.J. CFD Vision 2030 Study: A Path to Revolutionary Computational Aerosciences; Technical report; NASA: Washington, DC, USA, 2014. 13. Darwish, M.; Moukalled, F. The Finite Volume Method in Computational Fluid Dynamics: An Advanced Introduction with OpenFOAM® and Matlab®; Springer: Berlin/Heidelberg, Germany, 2021. 14. Zou, S.; Yuan, X.F.; Yang, X.; Yi, W.; Xu, X. An integrated lattice Boltzmann and ﬁnite volume method for the simulation of viscoelastic ﬂuid ﬂows. J. Non-Newton. Fluid Mech. 2014, 211, 99–113. 15. Karypis, G.; Kumar, V. A fast and high quality multilevel scheme for partitioning irregular graphs. SIAM J. Sci. Comput. 1998, 20, 359–392. 16. Gibbs, N.E.; Poole, W.G., Jr.; Stockmeyer, P.K. An algorithm for reducing the bandwidth and proﬁle of a sparse matrix. SIAM J. Numer. Anal. 1976, 13, 236–250. 17. Cuthill, E.; McKee, J. Reducing the bandwidth of sparse symmetric matrices. In Proceedings of the 1969 24th National Conference, New York, NY, USA, 26–28 August 1969; pp. 157–172. 18. Sloan, S. An algorithm for proﬁle and wavefront reduction of sparse matrices. Int. J. Numer. Methods Eng. 1986, 23, 239–251. 19. Akhras, G.; Dhatt, G. An automatic node relabelling scheme for minimizing a matrix or network bandwidth. Int. J. Numer. Methods Eng. 1976, 10, 787–797. 20. George, A.; Liu, J.W. Computer Solution of Large Sparse Positive Deﬁnite; Prentice Hall Professional Technical Reference: Hoboken, NJ, USA, 1981. 21. de Oliveira, S.G.; Silva, L.M. An ant colony hyperheuristic approach for matrix bandwidth reduction. Appl. Soft Comput. 2020, 94, 106434. 22. Pop, P.; Matei, O.; Comes, C.A. Reducing the bandwidth of a sparse matrix with a genetic algorithm. Optimization 2014, 63, 1851–1876. 23. Xiao-wei, G.; Chao, L.; Jie, L.; Chuan-fu, X.; Chun-ye, G.; Li-juan, C. A highly scalable general purpose CFD software architecture and its prototype implementation. Comput. Eng. Sci. 2020, 42, 2117–2124. 24. Xu, C.; Deng, X.; Zhang, L.; Fang, J.; Wang, G.; Jiang, Y.; Cao, W.; Che, Y.; Wang, Y.; Wang, Z.; et al. Collaborating CPU and GPU for large-scale high-order CFD simulations with complex grids on the TianHe-1A supercomputer. J. Comput. Phys. 2014, 278, 275–297. 25. Karypis, G.; Schloegel, K.; Kumar, V. Parmetis: Parallel Graph Partitioning and Sparse Matrix Ordering Library; University of Minnesota: Minneapolis, MN, USA, 1997. 26. Burgess, D.; Giles, M.B. Renumbering unstructured grids to improve the performance of codes on hierarchical memory machines. Adv. Eng. Softw. 1997, 28, 189–201. 27. Zhang, H.; Guo, X.W.; Li, C.; Liu, Q.; Xu, H.; Liu, J. Accelerating FVM-Based Parallel Fluid Simulations with Better Grid Renumbering Methods. Appl. Sci. 2022, 12, 7603. https://doi.org/10.3390/app12157603. 28. Azad, A.; Jacquelin, M.; Buluç, A.; Ng, E.G. The reverse Cuthill-McKee algorithm in distributed-memory. In Proceedings of the 2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS), Buena Vista, FL, USA, 29 May–2 June 2017; pp. 22–31. 29. Versteeg, H.K.; Malalasekera, W. An Introduction to Computational Fluid Dynamics: The Finite Volume Method; Pearson Education: London, UK, 2007.

Applied Sciences – Multidisciplinary Digital Publishing Institute

**Published: ** Oct 11, 2022

**Keywords: **mesh renumbering; thermal engineering hydraulics; CFD; YH-ACT; RCM; parallel calculation

Loading...

You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!

Read and print from thousands of top scholarly journals.

System error. Please try again!

Already have an account? Log in

Bookmark this article. You can see your Bookmarks on your DeepDyve Library.

To save an article, **log in** first, or **sign up** for a DeepDyve account if you don’t already have one.

Copy and paste the desired citation format or use the link below to download a file formatted for EndNote

Access the full text.

Sign up today, get DeepDyve free for 14 days.

All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.