Parallel scheduling based simulation scenario multi-objective calculation method and device, and server

Abstract

The application provides a simulation scene multi-target calculation method and device based on parallel scheduling and a server, relates to the technical field of industrial simulation application, and comprises the following steps: obtaining a parameter set of multiple simulation tasks, performing task clustering processing on the parameter set according to parameter similarity, and obtaining multiple task clusters; distributing each task cluster to different computing nodes, performing parallel solving processing of the simulation tasks on the task clusters in each computing node based on an adaptive optimization model of a message passing interface, and obtaining a local simulation result corresponding to each computing node; performing gathering and reorganization processing on the local simulation result, obtaining a complete simulation result dataset, performing compression and index storage processing on the simulation result dataset, generating a lightweight simulation result, performing stream visual processing on the lightweight simulation result, and obtaining a target interactive simulation visual result. The application can significantly improve the simulation efficiency of simulation scene multi-target calculation.

Description

Technical Field

[0001] This invention relates to the technical field of industrial simulation applications, and in particular to a method, apparatus and server for multi-objective computation of simulation scenarios based on parallel scheduling. Background Technology

[0002] Currently, cloud-based industrial simulation applications have become an important means of product design, performance verification, and optimization. With the increasing complexity of product structures and the growing demand for multiphysics coupling analysis, traditional single-machine simulation modes are no longer sufficient to meet the requirements of high precision and high efficiency. Related technologies suggest that existing cloud-based industrial simulation applications can be combined with distributed high-performance computing (HPLC) to provide parallel computing capabilities for large-scale simulations. However, HPLC tasks suffer from uneven load distribution across multiple nodes and improper MPI communication parameter configuration, resulting in low utilization of computing resources, long solution times, and large amounts of simulation results that are difficult to analyze in real time. Therefore, the above solutions still have significant shortcomings in automated parallel scheduling, communication latency control, result convergence consistency, and lightweight visualization, making it difficult to meet the comprehensive requirements of efficiency, intelligence, and real-time performance in multi-objective simulation computing. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method, apparatus and server for multi-objective computation in simulation scenarios based on parallel scheduling, which can significantly improve the simulation efficiency of multi-objective computation in simulation scenarios.

[0004] In a first aspect, embodiments of the present invention provide a multi-objective computation method for simulation scenarios based on parallel scheduling. The method includes: acquiring parameter sets of multiple simulation tasks, and performing task clustering processing on the parameter sets according to parameter similarity to obtain multiple task clusters; allocating each task cluster to different computing nodes based on a preset task scheduling mapping table, and performing parallel solution processing of simulation tasks on the task clusters within each computing node based on an adaptive tuning model of a message passing interface to obtain local simulation results corresponding to each computing node; performing aggregation and reorganization processing on the local simulation results to obtain a complete simulation result dataset, and performing compression and indexed storage processing on the simulation result dataset to generate lightweight simulation results, and performing streaming visualization processing on the lightweight simulation results to obtain interactive simulation visualization results of the targets.

[0005] In one implementation, the step of clustering parameter sets based on parameter similarity to obtain multiple task clusters includes: performing distance calculation on each parameter set using a preset parameter difference matrix to obtain the difference distance corresponding to each parameter set; dividing simulation tasks with difference distances less than a preset difference distance threshold into a unified task cluster, so that tasks within each task cluster can share preprocessing data and reduce repeated loading and initialization operations.

[0006] In one implementation, after obtaining the difference distances corresponding to each set of parameters, the method includes: determining parameter similarity using the difference distances; the smaller the difference distance between any two simulation tasks, the higher the similarity; and the larger the difference distance between any two simulation tasks, the lower the similarity.

[0007] In one implementation, the adaptive tuning model based on the message passing interface performs parallel solution processing of simulation tasks for task clusters within each computing node to obtain local simulation results corresponding to each computing node. The steps include: using an asynchronous non-blocking communication mechanism to perform information exchange between each simulation task, so that the information exchange process and the local computation process of each simulation task are executed in parallel to obtain local simulation results corresponding to each computing node.

[0008] In one implementation, an asynchronous non-blocking communication mechanism is used to perform information exchange between various simulation tasks, enabling the information exchange process to be executed in parallel with the local computation process of each simulation task to obtain the local simulation results corresponding to each computing node. The steps include: dynamically optimizing the underlying communication parameters of the information exchange interface based on an adaptive tuning model to match the hardware topology and network status of the current computing cluster and generating a target parameter combination; and performing parallel solution processing on the simulation tasks in each task cluster based on the target parameter combination to obtain the local simulation results.

[0009] In one implementation, the step of compressing and indexing the simulation result dataset to generate lightweight simulation results includes: performing differential encoding on the aggregated and recombined simulation result dataset to compress the simulation result dataset, and using a result aggregation tree model to establish a multi-channel index table for the compressed simulation result dataset; and using the compressed simulation result dataset and the multi-channel index table to construct lightweight simulation results.

[0010] In one implementation, the step of performing streaming visualization processing on lightweight simulation results to obtain target interactive simulation visualization results includes: dividing the lightweight simulation results into a preset number of data blocks, and performing real-time rendering processing on the data blocks according to the user's perspective and zoom distance, so as to perform rendering detail layer control on the corresponding data blocks according to the user's perspective and zoom distance, thereby obtaining target interactive simulation visualization results.

[0011] Secondly, embodiments of the present invention also provide a multi-objective computing device for simulation scenarios based on parallel scheduling. The device includes: a task clustering module, which acquires parameter sets of multiple simulation tasks and performs task clustering processing on the parameter sets according to parameter similarity to obtain multiple task clusters; a parallel solving module, which allocates each task cluster to different computing nodes based on a preset task scheduling mapping table, and performs parallel solving processing on the task clusters within each computing node based on an adaptive tuning model of a message passing interface to obtain local simulation results corresponding to each computing node; and a convergence indexing module, which performs convergence and reorganization processing on the local simulation results to obtain a complete simulation result dataset, and performs compression and indexing storage processing on the simulation result dataset to generate lightweight simulation results, and performs streaming visualization processing on the lightweight simulation results to obtain interactive simulation visualization results of the targets.

[0012] Thirdly, embodiments of the present invention also provide a server, including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement any of the methods provided in the first aspect.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement any of the methods provided in the first aspect.

[0014] The embodiments of the present invention bring the following beneficial effects:

[0015] This invention provides a method, apparatus, and server for multi-objective computation in simulation scenarios based on parallel scheduling. After acquiring parameter sets for multiple simulation tasks, the method performs task clustering based on parameter similarity to obtain multiple task clusters. Then, based on a preset task scheduling mapping table, each task cluster is assigned to different computing nodes. Using an adaptive optimization model based on a message passing interface, the simulation tasks are solved in parallel for each task cluster within each computing node, yielding local simulation results for each computing node. Finally, the local simulation results are aggregated and reorganized to obtain a complete simulation result dataset. This dataset is then compressed and indexed for storage, generating lightweight simulation results. These lightweight simulation results are then subjected to streaming visualization to obtain interactive simulation visualization results. This invention significantly improves the simulation efficiency of multi-objective computation in simulation scenarios.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a multi-objective computation method for simulation scenarios based on parallel scheduling, provided in an embodiment of the present invention;

[0020] Figure 2 A schematic diagram of bending moment simulation provided for an embodiment of the present invention;

[0021] Figure 3 A schematic diagram illustrating the specific process of a multi-objective computation method for simulation scenarios based on parallel scheduling, provided in an embodiment of the present invention;

[0022] Figure 4 A schematic diagram of an equivalent stress result provided in an embodiment of the present invention;

[0023] Figure 5 A schematic diagram of the structure of a multi-objective computing device for simulation scenarios based on parallel scheduling, provided in an embodiment of the present invention;

[0024] Figure 6 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Currently, cloud-based industrial simulation applications have become an important means for product design, performance verification, and optimization in modern industrial R&D. With the increasing complexity of product structures and the growing demand for multiphysics coupling analysis, traditional single-machine simulation modes are no longer sufficient to meet the requirements of high precision and high efficiency. Distributed high-performance computing (HPC) provides parallel computing capabilities for large-scale simulations, but existing cloud-based industrial simulation applications combined with HPC still face significant bottlenecks in solution efficiency and visualization response: on the one hand, uneven task load across multiple nodes and improper MPI communication parameter configuration lead to low utilization of computing resources and long solution times; on the other hand, the massive amount of simulation results results in a lag in the post-processing stage on the web, affecting the visualization and interactive experience and hindering real-time analysis.

[0027] Meanwhile, industrial simulation applications are gradually developing towards intelligence and multi-objective optimization. Many simulation scenarios based on parametric modeling require multiple rounds of iterative calculations, such as structural strength optimization and flow field tuning. These multi-parameter, multi-objective optimization tasks place higher demands on computing resource scheduling, communication synchronization, and visualization feedback. Traditional serial scheduling and static task allocation modes are difficult to adapt to the complex needs of dynamic load and parallel optimization.

[0028] Current industrial simulation applications primarily rely on manually setting MPI communication parameters and fixed task partitioning strategies, lacking adaptive optimization capabilities during the parallel solution phase. Manual tuning not only depends on experience but also struggles to handle dynamic changes in computational load across different simulation scenarios, often leading to uneven resource utilization and communication bottlenecks among nodes, thus reducing overall solution efficiency. Furthermore, existing communication mechanisms mostly employ blocking transmission, resulting in high cross-node latency and limiting the scalability of parallel computing.

[0029] In the post-processing stage, traditional solutions still rely on centralized result aggregation and single-field visualization. These methods cannot efficiently merge multi-node output data, easily leading to data redundancy and inconsistent field values. Furthermore, the overall rendering process is lagging, preventing real-time visualization or streaming updates of results. In addition, for multi-objective optimization simulations, existing solutions typically separate parameter iteration from the solution process, lacking a unified parallel scheduling and data reuse mechanism, resulting in low optimization efficiency.

[0030] Based on this, the simulation scenario multi-objective computing method, device and server based on parallel scheduling provided by the present invention can achieve end-to-end optimization of the solution and post-processing stages through MPI communication parameter tuning, task clustering scheduling, result aggregation and compression and streaming visualization rendering, which significantly improves the overall performance and interaction efficiency of distributed simulation computing.

[0031] See Figure 1The diagram shows a flowchart of a multi-objective computation method for a simulation scenario based on parallel scheduling. The method mainly includes the following steps S102 to S106:

[0032] Step S102: Obtain parameter sets for multiple simulation tasks, and perform task clustering processing on the parameter sets based on parameter similarity to obtain multiple task clusters.

[0033] In one implementation, before performing task clustering on the parameter set, simulation task generation and parameter definition are required to obtain the parameter set of the simulation task. In practical applications, taking passenger car wheel hub bending simulation as an example: the passenger car wheel hub bending simulation involves a fixed wheel hub and rim, with the contact surface of the wheel hub bolt holes subjected to bending moment loads. To explore the load-bearing capacity of the bending moment at various angles of the wheel hub, an iterative calculation method based on angle is introduced in the simulation application, starting from the initial zero-degree position and ending at the position of a complete 360-degree circle. During the iterative calculation process, to ensure the accuracy of the results, the angle of the bending moment on the wheel hub is adjusted for iterative calculation to complete the calculation of the complete circumference. In the actual passenger car wheel hub bending simulation application, a remote force component is introduced to achieve the effect of an equivalent loading arm. In the original simulation process, a bending moment around a single axis (taking the x-axis as an example) is applied to a remote point. However, to reproduce the real passenger car bending test, a formula-driven method is used to reproduce the influence of the circumferential moment. In one implementation, a sine function is applied to the original x-axis moment, specifically expressed as:

[0034]

[0035] in, The magnitude of the bending moment about the x-axis. The test standard bending moment for the modified wheel hub model is given by time, which is the iteration time in each iteration step during the simulation solution.

[0036] Furthermore, the standard bending moment for the wheel hub test was set at 4135 N·m, with a total of 60 simulation iterations. Each simulation iteration completed the calculation from 0 to 1 second. For the bending moment performance after formula-driven calculation, please refer to [link to relevant documentation]. Figure 2 The diagram illustrates a specific flowchart of a multi-objective computation method for simulation scenarios based on parallel scheduling (the horizontal axis represents the number of simulation iterations, and the vertical axis represents the corresponding bending moment vector value). To achieve the effect of applying bending moment along a circle, a single-axis torque drive is insufficient. Therefore, a cosine function bending moment drive, orthogonal to the x-axis, is simultaneously introduced along the y-axis. The specific formula is as follows:

[0037]

[0038] in, This represents the magnitude of the bending moment about the y-axis.

[0039] By superimposing bending moment vectors in the orthogonal x-axis and y-axis directions, a complete 360° bending simulation can be completed, and the result frame with the largest equivalent stress can be locked to show the weak point of the wheel hub.

[0040] To verify the effectiveness of this invention, passenger car wheel hub bending evaluation simulation is selected as a typical industrial simulation application scenario in this embodiment. The communication layer adopts Intel MPI 2021.8, and the computing tasks run on a distributed high-performance computing cluster equipped with 56 cores × 10 nodes (a total of 560 cores). The entire simulation process, from front-end task generation and scheduling execution to result aggregation and visualization output, constructs a complete end-to-end multi-objective computing and optimization technology route, realizing the automation, parallelization, and efficiency of the computing process.

[0041] During the simulation task generation and packaging phase, users first set the wheel hub's geometric model, material properties, boundary constraints, and bending load parameters through the platform's parametric configuration interface. Based on preset angle stepping rules, the system automatically loads sine and cosine driving functions to achieve periodic changes in bending moment along the x and y axes, thereby accurately simulating the actual bending load process of the wheel hub under circumferential stress.

[0042] Subsequently, the system automatically generates 60 sets of working condition input files with different combinations of angles and load amplitudes, and records the parameter sets and control commands for each working condition in JSON (JavaScript Object Notation) format. The generated preprocessing files and parameter configurations are uniformly packaged into a task package, which includes geometric files, load definitions, simulation control scripts, and job description files, i.e., the parameter set of the aforementioned multiple simulation tasks.

[0043] Step S104: Based on the preset task scheduling mapping table, each task cluster is assigned to different computing nodes, and based on the adaptive tuning model of the message passing interface, the simulation tasks are solved in parallel for each task cluster within each computing node to obtain the local simulation results corresponding to each computing node.

[0044] In one implementation, based on a preset task scheduling mapping table, task clusters are allocated to different computing nodes to achieve task scheduling. During the parallel scheduling phase, the system automatically generates a task scheduling mapping table, created at runtime by a containerized script, to maintain the binding relationship between task clusters and computing nodes, ensuring the traceability and controllability of the computing process. When multi-node jobs are submitted to the cluster, the master node automatically generates a hostfile file based on the number of available cores on each node, realizing on-demand dynamic allocation of computing resources and multi-level parallel scheduling, thereby maximizing cluster computing efficiency while maintaining data consistency.

[0045] After the task package is uploaded to the job management and scheduling layer, the system automatically completes job registration, status monitoring, resource application, and statistical information collection. After the task is submitted, the task organization and scheduling module generates a multi-condition task index table based on the returned job number, and uses a hash mapping mechanism to establish a mapping relationship between the condition number and the high-performance computing node, realizing efficient binding and rapid allocation of conditions and computing resources, thereby providing a structured input and dynamic scheduling foundation for the subsequent parallel solution stage.

[0046] In another implementation, during the parallel scheduling and MPI task distribution phase, the system invokes the distributed task distribution mechanism of Intel MPI (Message Passing Interface, a high-performance distributed communication library provided by Intel). It achieves dual-layer parallelism at both the task-level and process-level through the `mpirun` command (the MPI job startup command used to initiate multi-process parallel computing) and a multi-level process mapping strategy. At the task-level scheduling layer, the system clusters tasks based on parameter similarity, and each task cluster is assigned to an independent computing node, achieving task-level parallelism. At the process-level scheduling layer, each node undergoes further fine-grained parallel partitioning, and the CPU binding strategy is optimized using the `I_MPI_PIN_DOMAIN` parameter (a parameter controlling the CPU affinity setting of MPI processes to reduce inter-core communication latency) to reduce inter-process communication latency. During communication, the automatic detection function of Intel MPI 2021.8's FabricProvider (network interconnection driver layer interface, responsible for the implementation of MPI communication at the underlying network) (here specified as `ofi`, i.e., OpenFabricsInterfaces, used to support high-speed interconnection communication) is utilized to achieve high-speed interconnection optimization between nodes. Meanwhile, by adjusting communication parameters such as I_MPI_ADJUST_BCAST=3 (a parameter that controls the MPI broadcast algorithm type, with a value of 3 indicating the use of a hierarchical tree-based broadcast acceleration algorithm), broadcasting and synchronization operations are accelerated, thereby further improving the overall parallel efficiency.

[0047] Step S106: Aggregate and reorganize the local simulation results to obtain a complete simulation result dataset. Compress and index the simulation result dataset for storage to generate lightweight simulation results. Perform streaming visualization processing on the lightweight simulation results to obtain the target interactive simulation visualization results.

[0048] In one implementation, during the front-end visualization and reduced-order analysis phase, after receiving the returned data, the system calls the POD algorithm (Proper Orthogonal Decomposition, a dimensionality reduction technique for extracting main feature modes) to extract the main feature patterns, thereby achieving a reduced-order display of the simulation results. The front-end employs streaming rendering technology based on LOD (Level of Detail Control), dynamically loading corresponding block data according to the viewpoint and node labels to achieve rapid interactive visualization of large-scale simulation results. Users can view the stress distribution trends and the evolution of weak areas from different angles in real time on the browser. That is, when viewing the front of the model, the data on the back is not loaded. LOD layered detail control means that when viewing the overall picture in a thumbnail, the system uses a coarser model (lower precision) to render each data block, quickly drawing the outline; when zooming in to view the details on the wheel hub, the system switches to a more refined model (high precision) to render the data block containing the hole, allowing users to clearly see every texture and improving the intuitiveness and analysis efficiency of the simulation results.

[0049] The above-mentioned simulation scenario multi-objective computation method based on parallel scheduling provided in this embodiment of the invention first generates a set of working conditions during front-end parameter configuration, submits task scheduling through a Docker-encapsulated post-processing container, generates task clusters based on parameter similarity, and establishes multi-level parallel scheduling at the MPI level. After the solution is completed, the multi-node output field data is uniformly reorganized and indexed through result aggregation, and finally, the Web-based display module realizes block visualization and reduced-order output. Without changing the core algorithm of the solver, end-to-end performance optimization of the solution and post-processing stages is achieved through key technologies such as MPI communication parameter optimization, task clustering scheduling, result aggregation compression, and streaming visualization rendering, which significantly improves the simulation efficiency of multi-objective computation in simulation scenarios.

[0050] See Figure 3 The diagram shows a specific flowchart of a multi-objective computation method for simulation scenarios based on parallel scheduling. This embodiment of the invention also provides an implementation method for multi-objective computation in simulation scenarios, as detailed in (1) to (4) below:

[0051] (1) Preprocessing: The distance calculation of each parameter set is performed by the preset parameter difference matrix to obtain the difference distance corresponding to each parameter set. The simulation tasks with difference distance less than the preset difference distance threshold are divided into a unified task cluster so that the tasks in each task cluster can share the preprocessing data and reduce repeated loading and initialization operations. Among them, the difference distance can be used to determine the similarity of parameters. When using the difference distance to determine the similarity of parameters, the smaller the difference distance between any two simulation tasks, the higher the similarity; the larger the difference distance between any two simulation tasks, the lower the similarity.

[0052] In one implementation, during the parameterized organization and clustering scheduling of simulation tasks: In the task generation and allocation phase, to further improve the parallelism of task scheduling and data reuse rate, a parameterized task organization and clustering scheduling mechanism is introduced during simulation initialization. This mechanism establishes a unified parameterized description model to achieve intelligent clustering and efficient resource mapping of multi-condition simulation tasks.

[0053] First, the system defines the set of all tasks to be computed as follows:

[0054]

[0055] in, This represents the k-th parameter of the i-th task.

[0056] To avoid repeatedly loading similar models, this invention uses a parameter difference distance matrix:

[0057]

[0058] in, The value of task i on the k-th parameter (which can be a continuous type such as stress intensity, or a discrete / categorical type parameter after being encoded). This refers to the total number of parameters (dimensions), such as load magnitude, material modulus, boundary condition on / off flags, etc. The normalization scale for the k-th parameter is the sample standard deviation, which is used to put parameters of different dimensions and numerical ranges on the same order of magnitude to avoid a parameter of a large dimension dominating the distance. The weights (non-negative real numbers) of the k-th parameter satisfy normalization. .

[0059] For the two simulation tasks i and j, calculate the normalized sum of squares of the differences along the parameter dimensions k=1 to m, and then sort them by weight. We obtain a scalar distance by weighting and taking the square root. The smaller the distance, the more similar the two tasks are in the parameter space, and the more suitable they are to be placed in the same cluster. Each task cluster only needs to perform geometry and mesh initialization operations once, and tasks within the cluster share preloaded data, thereby significantly reducing I / O operations and preprocessing overhead in large-scale simulations.

[0060] (2) MPI optimization solution: An asynchronous non-blocking communication mechanism is adopted to perform information exchange between various simulation tasks, so that the information exchange process and the local calculation process of each simulation task are executed in parallel to obtain the local simulation results corresponding to each computing node, thereby solving the problem of data transmission blocking during communication. The parameter configuration problem in this mode can be solved by using an adaptive tuning model. That is, based on the adaptive tuning model, the underlying communication parameters of the information exchange interface are dynamically optimized to match the hardware topology and network status of the current computing cluster and generate target parameter combinations. Then, based on the target parameter combinations, the simulation tasks in each task cluster are solved in parallel to obtain local simulation results.

[0061] In one implementation, during the computational solution phase, each MPI subprocess (MPI RankProcess), representing an independent process unit in distributed parallel computing, executes the hub bending solution at the corresponding angle. To further improve the efficiency of communication and computational parallelism, an asynchronous non-blocking communication mechanism can be adopted. That is, MPI_Isend (asynchronous send function) and MPI_Irecv (asynchronous receive function) are used to achieve parallel execution of data exchange and local computation, effectively reducing communication waiting time. This is equivalent to the existing scheme where process A sends data to process B and needs to wait for process B's reply before proceeding to the next step of computation, but now process A can start computation directly after sending the data.

[0062] Meanwhile, an adaptive tuning mechanism based on the mpitune tool (an automatic tuning tool provided by Intel MPI for optimizing MPI communication parameters under different hardware and network topologies) is introduced to dynamically optimize key MPI parameters I_MPI_ADJUST_ALLTOALL (a parameter that controls the selection of the all-to-all communication algorithm) and I_MPI_SPIN_COUNT (a parameter that controls the number of busy-wait loops to balance CPU utilization and latency). This significantly improves both CPU utilization and inter-node communication efficiency during the parallel solution phase. In other words, automatic tuning avoids manually setting these complex low-level parameters. Instead, it uses intelligent tools to automatically test on actual computing clusters and find the optimal parameter combination that best suits the current hardware and network.

[0063] Specifically, after each node completes the solution, the result aggregation and indexing module is started. Through MPI aggregation communication operations, MPI_Gather (aggregation collection function, used to collect data from multiple processes to the main process) and MPI_Reduce (reduction function, used to perform aggregation operations such as summation and averaging between multiple processes) are used to uniformly collect and summarize the stress field and displacement field data output by multiple nodes.

[0064] The core logic of MPI parallel solution and communication optimization aims to maximize the utilization of distributed computing resources. Through a multi-layer parallel architecture and adaptive communication strategy, it achieves efficient solution and dynamic load balancing for multi-condition simulation.

[0065] This invention employs a two-layer parallel structure in the solution phase. The outer layer parallelism involves multi-condition allocation at the task level, with each condition executed as an independent solution instance in parallel. The inner layer parallelism utilizes MPI domain decomposition technology for the single-condition simulation domain, achieving fine-grained parallel computation.

[0066] 1. Task-level scheduling logic: For a task set T, the master node (rank0) executes a cyclic allocation algorithm i modulo R based on the number of nodes R to ensure a balanced distribution of tasks among the nodes. During job execution, the system monitors the CPU utilization, network utilization, and I / O latency of each node in real time. This is achieved by defining a node performance margin function:

[0067]

[0068] For the first The current CPU utilization of a node represents the node's computational intensity. For the first The throughput capacity of each node; For the first Network utilization of individual nodes; , , This represents the average metric for all nodes in the cluster at that moment. , , For the weighting coefficients, satisfying + + =1.

[0069] when When the load is less than the set threshold ε, the system considers the node's performance margin insufficient. The master node triggers a task reallocation mechanism through MPI communication to migrate the remaining subtasks to nodes with higher performance margins, thereby achieving real-time load balancing.

[0070] in, , , Configurable weighting coefficients are used to balance the importance of CPU computational load, I / O bandwidth, and network communication. This mechanism significantly reduces the overall solution time fluctuation caused by differences in node performance, improving the scalability and stability of parallel simulation.

[0071] 2. Domain Decomposition and Communication Model: In the parallel optimization process of this invention, to quantitatively analyze the performance variation of the distributed solver under different numbers of processes, parallel speedup, parallel efficiency, and total communication overhead are introduced to theoretically model and verify the overall performance. Within each simulation task, the solution domain is divided into... Each subdomain is processed by one MPI process. The parallel speedup is defined as:

[0072]

[0073] Where p is the number of processes participating in parallelism; f is the proportion of the parallelizable portion (0 ≤ f ≤ 1); that is, the proportion of the total serial time that can be parallelized. The corresponding serial portion proportion is 1. f.

[0074] Parallel efficiency is:

[0075]

[0076] The theoretical model shows that when f≈1 and the communication cost is low, the system approaches ideal acceleration.

[0077] In practical applications, communication latency cannot be ignored. The total communication overhead can be expressed as:

[0078]

[0079] Where α is the message initiation delay, β is the transmission time constant per byte, and m is the message size. For the number of communications. To reduce The present invention employs the following strategy:

[0080] Non-blocking communication: During the parallel solution phase, the traditional blocking communication operations MPI_Send / MPI_Recv (basic blocking communication operations for sending and receiving messages) are replaced with MPI_Isend / MPI_Irecv (non-blocking communication operations for asynchronous sending and receiving), and used in conjunction with MPI_Waitall (a synchronous operation that waits for all asynchronous communication operations to complete), thereby achieving overlap between the computation and communication processes (i.e., continuing to execute computational tasks during data transmission) to improve parallel efficiency.

[0081] Layered buffering strategy: During intra-node communication, the shared memory channel is used first, so that inter-process data transmission within the same computing node bypasses the network layer and is completed directly through the shared memory area; while during cross-node communication, the optimal transmission path is automatically selected based on the OFI (Open Fabrics Interfaces) communication layer to achieve high-bandwidth, low-latency data exchange.

[0082] Adaptive parameter tuning: Using the `mpitune_fast` tool provided by Intel MPI (a lightweight tuner for automatically generating optimal communication parameter combinations), the optimal configuration parameter set is automatically generated based on the actual cluster architecture and communication mode. The following parameters are used in this invention: `I_MPI_FABRICS` (specifying the communication layer structure as `ofi`), used to define the underlying communication mechanism for inter-process data transmission in MPI; `I_MPI_EAGER_THRESHOLD` (controlling the Eager protocol packet size threshold in bytes), used to balance the latency of small message transmissions with the bandwidth utilization of large messages; and `I_MPI_PIN_DOMAIN` (defining the field of the process binding strategy, specified as `socket`), used to optimize process-memory affinity under multi-core CPUs. After tuning, the system caches the optimal parameter combination in a local configuration file for reuse in subsequent job submissions, reducing the overhead of repeated parameter tuning.

[0083] When inter-process communication and computation overlap in MPI, the communication is completely hidden in the computation stage, thus achieving a near-linear speedup.

[0084] A key characteristic of this type of industrial simulation application is the need for large-scale iterative calculations. This example of an industrial simulation application effectively utilizes large-scale computational operations. See [link to relevant documentation]. Figure 4 The diagram shown is a schematic representation of the equivalent stress results. It displays all equivalent stress results after iteration. The horizontal axis represents the number of the 60 simulations. The upper curve represents the maximum equivalent stress result for each operation, the middle curve represents the average equivalent stress result for each operation, and the lower curve represents the minimum equivalent stress result for each operation.

[0085] (3) Results aggregation and indexing: Differential encoding is performed on the aggregated and recombined simulation results dataset to compress the simulation results dataset. A multi-channel index table of the compressed simulation results dataset is established using the results aggregation tree model. Then, lightweight simulation results are constructed using the compressed simulation results dataset and the multi-channel index table.

[0086] In one implementation, after the results are aggregated, the system invokes a built-in lightweight post-processing container to perform data compression and transmission preparation. To achieve efficient data transmission and remote display, a strategy based on differential coding (a compression algorithm for storing changes in adjacent data) and the ZFP algorithm (ZFloating-Point Compression, a lossy compression algorithm for scientific computing with a high compression ratio) is adopted to compress the simulation field data to 20% to 30% of its original volume. After compression, the data is transmitted back to the front-end visualization service node through a high-performance transmission channel based on a gRPC (Google Remote Procedure Call, a framework for efficient bidirectional streaming communication) streaming interface, ensuring low latency and high reliability during the data transmission phase.

[0087] Further details regarding result aggregation and multi-channel indexed storage include: After simulation, each MPI process generates a local result file. To reduce the burden of manual merging, this invention employs a result aggregation tree structure. The master node, rooted at rank0, aggregates field data through multi-level MPI_Gather and MPI_Reduce operations. To avoid node number misalignment, the system uses coordinate distance-based matching logic to remap nodes. After aggregation, the system automatically generates a result index table.

[0088] All results are divided into two categories: 1. The main channel stores key physical quantities (such as displacement, velocity, and pressure field) and outputs them in VTK format; 2. The auxiliary channel stores derived variables (such as equivalent stress and energy density) and saves them in HDF5 format. This multi-channel structure facilitates front-end display and also supports subsequent order reduction modeling.

[0089] (4) Visualization and reduction of scale: The lightweight simulation results are divided into a preset number of data blocks, and the data blocks are rendered in real time according to the user's perspective and zoom distance. The rendering details of the corresponding data blocks are controlled according to the user's perspective and zoom distance to obtain the target interactive simulation visualization results.

[0090] In one implementation, after the MPI optimization solution is completed, the multi-node simulation results can be processed for streaming visualization and reduced-order feature extraction, achieving dual optimization of data transmission and display performance: First, during the visualization data loading stage, the system divides the overall computational domain into multiple independent data blocks: Each block corresponds to a spatial region and its set of field variables. The backend transmits the corresponding blocks as needed based on the frontend camera's view frustum, achieving dynamic loading and updating at the block level, thereby reducing the memory overhead and network bandwidth consumption of one-time loading. During the rendering phase, the system defines layered detail control functions based on camera distance and viewing angle:

[0091]

[0092] in, , , These represent high-precision, medium-precision, and low-precision rendering levels, respectively. This level function allows the system to adjust the rendering precision in real time based on the user's viewpoint and zoom distance during interaction, forming a distance-aware LOD strategy. When the viewing distance is far, the system only loads low-resolution geometric contours; while when the camera gets closer, it automatically switches to a high-resolution model to ensure local accuracy and smooth interaction.

[0093] In the data reduction section, the system uses the Singular Value Decomposition (SVD) method to optimize the converged field data matrix. Decompose:

[0094]

[0095] Among them, matrix Characterize the principal modal basis of the node space. Let V be the singular value matrix arranged in descending order of energy, and V be the temporal or load eigenvector of the state space. By truncating the first k singular values ​​and their corresponding modes, a low-dimensional space representing the dominant physical characteristics can be obtained. This result can be further input into the surrogate model to achieve rapid reconstruction of simulation results and surrogate prediction.

[0096] The beneficial effects of this module are as follows: through a three-level optimization strategy of block division, LOD and dimensionality reduction, it achieves a reduction of more than 60% in data transmission bandwidth, an increase of about 2 times in front-end visualization frame rate, and a reduction of about 80% in model calculation time after dimensionality reduction, effectively improving the overall response speed and scalability of the high-performance simulation system in the result display and AI training stages.

[0097] Furthermore, to reduce the uncertainty caused by manual parameter tuning and improve the performance stability of the communication layer, this invention introduces a communication adaptive tuning mechanism in the MPI optimization solution module. This mechanism, based on the external automatic tuning tool mpitune_fast provided by Intel MPI, can automatically explore combinations of communication parameters under different hardware topologies and load characteristics to form a targeted optimal configuration. Unlike the MPI tuning during runtime in the aforementioned solution phase, this part is tuned during the pre-execution process.

[0098] Specifically, during the optimization phase, the system executes the following instructions:

[0099] mpitune_fast-app"xxxx"-otune_out

[0100] The `app` parameter specifies the simulation execution program and input file, while `-o` defines the output path for the tuning results. This process is automatically completed by the parameter search engine within `mpitune_fast`, systematically evaluating the combination space of typical variables, including:

[0101] Communication structure layer selection (I_MPI_FABRICS=shm:ofi,shm:tcp); transport layer provider (FI_PROVIDER=mlx,psm3,tcp); process binding method (I_MPI_PIN_DOMAIN=core,socket); message threshold parameters (I_MPI_EAGER_THRESHOLD, etc.).

[0102] The system ultimately outputs the configuration file mpitune.conf corresponding to the optimal parameter set, which is saved in the output directory tune_out.

[0103] During runtime, the main control script can directly reuse the optimization results by calling the command: mpirun-configtune_out. This command loads the communication parameters saved in the configuration file during MPI initialization, enabling the optimal configuration obtained during the tuning phase to be reproduced in subsequent multi-task computations, thus achieving cross-task migration and reuse of communication layer performance.

[0104] In the passenger car wheel hub bending simulation application, the post-processing needs to display sixty simulation results simultaneously, and display the circumferential simulation stress and deformation cloud map through multi-frame playback. The file size of the results is as large as 27G. The simulation results can be loaded and played smoothly on the web through the streaming method, and the interaction is guaranteed to be smooth.

[0105] In summary, the present invention has the following significant advantages:

[0106] 1. Automated Parallel Scheduling: Eliminates dependence on manual parameter combinations through task clustering and MPI adaptive tuning mechanisms; 2. Minimized Communication Overhead: Effectively reduces cross-node latency based on non-blocking communication and Fabric automatic selection mechanism; 3. Efficient Result Aggregation: Ensures field data consistency and aggregation speed by adopting hierarchical merging algorithm and KD-Tree remapping mechanism; 4. Streaming Visualization and Reduced-Order Output: Enables lightweight display of simulation results and reuse of AI training data; 5. High Versatility: Applicable to various MPI solution scenarios such as CFD, structural mechanics, and electromagnetic simulation.

[0107] Regarding the multi-objective computation method for simulation scenarios based on parallel scheduling provided in the foregoing embodiments, this invention provides a multi-objective computation device for simulation scenarios based on parallel scheduling, see [link to previous document]. Figure 5The diagram shows a multi-objective computing device for a simulation scenario based on parallel scheduling. The device includes the following parts:

[0108] The task clustering module 502 obtains parameter sets of multiple simulation tasks and performs task clustering processing on the parameter sets based on parameter similarity to obtain multiple task clusters;

[0109] The parallel solving module 504, based on a preset task scheduling mapping table, allocates various task clusters to different computing nodes, and performs parallel solving of simulation tasks on the task clusters within each computing node based on the adaptive tuning model of the message passing interface, so as to obtain the local simulation results corresponding to each computing node.

[0110] The aggregation index module 506 aggregates and reorganizes local simulation results to obtain a complete simulation result dataset. It then compresses and indexes the simulation result dataset for storage, generating lightweight simulation results. These lightweight simulation results are then used for streaming visualization to obtain the target interactive simulation visualization results.

[0111] The simulation scenario multi-objective computing device based on parallel scheduling provided in this application embodiment can significantly improve the simulation efficiency of multi-objective computing in simulation scenarios.

[0112] In one embodiment, when performing task clustering processing on parameter sets based on parameter similarity to obtain multiple task clusters, the task clustering module 502 is further configured to: perform distance calculation processing on each parameter set through a preset parameter difference matrix to obtain the difference distance corresponding to each parameter set; divide simulation tasks with difference distances less than a preset difference distance threshold into unified task clusters so that tasks within each task cluster can share preprocessing data, reducing repeated loading and initialization operations.

[0113] In one embodiment, after obtaining the difference distances corresponding to each parameter set, the task clustering module 502 is further configured to: determine parameter similarity using the difference distances; the smaller the difference distance between any two simulation tasks, the higher the similarity; the larger the difference distance between any two simulation tasks, the lower the similarity.

[0114] In one embodiment, when performing the adaptive tuning model based on the message passing interface, and performing parallel solution processing of simulation tasks for task clusters within each computing node to obtain the local simulation results corresponding to each computing node, the aforementioned parallel solution module 504 is further configured to: employ an asynchronous non-blocking communication mechanism to perform information exchange between each simulation task, so that the information exchange process and the local computation process of each simulation task are executed in parallel, so as to obtain the local simulation results corresponding to each computing node.

[0115] In one implementation, when performing the step of using an asynchronous non-blocking communication mechanism to execute information exchange between various simulation tasks, so that the information exchange process and the local computation process of each simulation task are executed in parallel to obtain the local simulation results corresponding to each computing node, the aforementioned parallel solving module 504 is further configured to: dynamically optimize the underlying communication parameters of the information exchange interface based on an adaptive tuning model to match the hardware topology and network status of the current computing cluster and generate a target parameter combination; and based on the target parameter combination, perform parallel solving on the simulation tasks in each task cluster to obtain the local simulation results.

[0116] In one embodiment, when performing the steps of compressing and indexing the simulation result dataset to generate lightweight simulation results, the above-mentioned aggregation index module 506 is further used to: perform differential encoding processing on the aggregated and recombined simulation result dataset to compress the simulation result dataset, and use the result aggregation tree model to establish a multi-channel index table of the compressed simulation result dataset; and use the compressed simulation result dataset and the multi-channel index table to construct lightweight simulation results.

[0117] In one embodiment, when performing the step of streaming visualization processing of lightweight simulation results to obtain target interactive simulation visualization results, the above-mentioned aggregation index module 506 is further used to: divide the lightweight simulation results into a preset number of data blocks, and perform real-time rendering processing on the data blocks according to the user's perspective and zoom distance, so as to perform rendering detail layer control on the corresponding data blocks according to the user's perspective and zoom distance, and obtain target interactive simulation visualization results.

[0118] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0119] This invention provides a server, specifically, the server includes a processor and a storage device; the storage device stores a computer program, which, when run by the processor, executes the method described in any of the above embodiments.

[0120] Figure 6 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. The server 100 includes: a processor 60, a memory 61, a bus 62, and a communication interface 63. The processor 60, the communication interface 63, and the memory 61 are connected through the bus 62. The processor 60 is used to execute executable modules, such as computer programs, stored in the memory 61.

[0121] The memory 61 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 63 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0122] Bus 62 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0123] The memory 61 is used to store programs. After receiving an execution instruction, the processor 60 executes the program. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 60 or implemented by the processor 60.

[0124] Processor 60 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 60 or by instructions in software form. Processor 60 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 61. Processor 60 reads the information in memory 61 and, in conjunction with its hardware, completes the steps of the above method.

[0125] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0126] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A parallel-schedule-based multi-objective computing method for simulation scenarios, characterized in that, The method includes: Obtain parameter sets for multiple simulation tasks, and perform task clustering processing on the parameter sets based on parameter similarity to obtain multiple task clusters; Based on a preset task scheduling mapping table, each task cluster is assigned to a different computing node, and based on an adaptive tuning model of the message passing interface, the task clusters within each computing node are processed in parallel to solve the simulation tasks, thereby obtaining the local simulation results corresponding to each computing node. The local simulation results are aggregated and reorganized to obtain a complete simulation result dataset. The simulation result dataset is then compressed and indexed for storage to generate lightweight simulation results. These lightweight simulation results are then subjected to streaming visualization to obtain the target interactive simulation visualization results. The step of clustering the parameter set according to parameter similarity to obtain multiple task clusters includes: performing distance calculation on each parameter set through a preset parameter difference matrix to obtain the difference distance corresponding to each parameter set; dividing the simulation tasks whose difference distance is less than a preset difference distance threshold into a unified task cluster, so that the tasks within each task cluster can share preprocessing data and reduce repeated loading and initialization operations; The step after obtaining the difference distances corresponding to each of the parameter sets includes: determining parameter similarity using the difference distances; the smaller the difference distance between any two simulation tasks, the higher the similarity; the larger the difference distance between any two simulation tasks, the lower the similarity.

2. The parallel-schedule-based multi-objective computation method for simulation scenarios according to claim 1, wherein, The adaptive tuning model based on the message passing interface performs parallel solution processing of simulation tasks for the task clusters within each computing node to obtain the local simulation results corresponding to each computing node, including: An asynchronous non-blocking communication mechanism is adopted to perform information exchange between the simulation tasks, so that the information exchange process and the local calculation process of each simulation task are executed in parallel to obtain the local simulation results corresponding to each computing node.

3. The parallel-schedule-based simulation-scenario multi-objective computation method of claim 2, wherein, The step of employing an asynchronous non-blocking communication mechanism to perform information exchange between the simulation tasks, enabling the information exchange process to be executed in parallel with the local computation process of each simulation task, in order to obtain the local simulation results corresponding to each computing node, includes: Based on the adaptive tuning model, the underlying communication parameters of the information exchange interface are dynamically optimized to match the hardware topology and network status of the current computing cluster and generate a target parameter combination. Based on the target parameter combination, the simulation tasks in each of the task clusters are solved in parallel to obtain the local simulation results.

4. The parallel-schedule-based simulation-scenario multi-objective computation method of claim 1, wherein, The step of compressing and indexing the simulation result dataset to generate lightweight simulation results includes: Differential encoding is performed on the aggregated and recombined simulation result dataset to compress the simulation result dataset, and a multi-channel index table of the compressed simulation result dataset is established using the result aggregation tree model; The lightweight simulation results are constructed using the compressed simulation result dataset and the multi-channel index table.

5. The parallel-schedule-based multi-objective computation method for simulation scenarios according to claim 1, wherein, The step of performing streaming visualization processing on the lightweight simulation results to obtain the target interactive simulation visualization results includes: The lightweight simulation result is divided into a preset number of data blocks, and the data blocks are rendered in real time according to the user's perspective and zoom distance. Based on the user's perspective and zoom distance, the corresponding data blocks are subjected to rendering detail layer control to obtain the target interactive simulation visualization result.

6. A parallel schedule based simulation scenario multi-objective computing apparatus, characterized by, The device includes: The task clustering module obtains parameter sets of multiple simulation tasks and performs task clustering processing on the parameter sets based on parameter similarity to obtain multiple task clusters; The parallel solution module, based on a preset task scheduling mapping table, allocates each task cluster to different computing nodes, and performs parallel solution processing on each task cluster within each computing node based on an adaptive tuning model of the message passing interface, to obtain the local simulation results corresponding to each computing node. The aggregation and indexing module aggregates and reorganizes the local simulation results to obtain a complete simulation result dataset. It then compresses and indexes the simulation result dataset for storage, generating lightweight simulation results. These lightweight simulation results are then subjected to streaming visualization to obtain the target interactive simulation visualization results. The step of clustering the parameter set according to parameter similarity to obtain multiple task clusters includes: performing distance calculation on each parameter set through a preset parameter difference matrix to obtain the difference distance corresponding to each parameter set; dividing the simulation tasks whose difference distance is less than a preset difference distance threshold into a unified task cluster, so that the tasks within each task cluster can share preprocessing data and reduce repeated loading and initialization operations; The step after obtaining the difference distances corresponding to each of the parameter sets includes: determining parameter similarity using the difference distances; the smaller the difference distance between any two simulation tasks, the higher the similarity; the larger the difference distance between any two simulation tasks, the lower the similarity.

7. A server, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in any one of claims 1 to 5.

Images