Distributed intelligent parallel deduction engine elastic resource arrangement method and system

Abstract

This invention discloses a method and system for elastic resource orchestration in a distributed intelligent parallel simulation engine. The method includes the following steps: dynamically marking time periods based on the predicted load state parameters of each simulation process in each subsequent simulation period, obtained from the dynamic load prediction of simulation data generated by each simulation process within each simulation period in the distributed simulation simulation system; dynamically setting and optimizing scheduling instructions based on the dynamic time period marking results and simulation event rules, and executing the dynamically set scheduling instructions after sorting them based on the topological constraint characteristics of the distributed simulation simulation system; and dynamically handling faults based on the container status parameters of each simulation container during the simulation process, and based on the container status assessment results and fault types. Through dynamic load prediction and intelligent scheduling, the periodic peak resource demand in distributed simulation is effectively smoothed, significantly improving overall resource utilization.

Description

Technical Field

[0001] This invention relates to the field of elastic resource data management technology, and in particular to an elastic resource orchestration method and system for a distributed intelligent parallel inference engine. Background Technology

[0002] In distributed simulation and deduction systems, the scenarios are becoming increasingly complex, the number of entities is vast, and interactions are frequent, leading to significant dynamic fluctuations in computational load. Traditional static resource allocation models struggle to cope with these alternating peak and trough load changes, easily causing resource bottlenecks or large amounts of idle resources. The elastic resource orchestration technology of distributed intelligent parallel deduction engines was developed precisely to address this core challenge. By monitoring the deduction load in real time and employing intelligent decision-making algorithms, it dynamically allocates and scales underlying computing resources on demand. This significantly improves cluster resource utilization efficiency while ensuring the real-time performance and continuity of the deduction, enabling modern large-scale simulation and deduction applications with high concurrency, low latency, and high availability.

[0003] Existing elastic resource orchestration methods primarily rely on monitors to continuously collect performance metrics of clusters and applications, analyze this time-series data using analyzers to determine the current state, calculate scaling decisions based on pre-defined elasticity strategies, and finally execute dynamic resource allocation by executors calling interfaces of the underlying infrastructure. While this technology is evolving from reactive scaling based on static thresholds to predictive scaling based on machine learning, aiming to improve resource efficiency, it still faces significant challenges when handling specific workloads such as distributed simulations.

[0004] For example, the invention patent announcement CN112860370B discloses a virtual machine resource allocation method, apparatus, system, and storage medium, which includes: a Network Functions Virtualization Orchestrator (NFVO) calculating the resources required by a Virtual Functions Unit (VNF); the NFVO querying the free resources of a Virtualization Infrastructure Manager (VIM); the NFVO determining the reservable resources of a VNF based on the free resources of the VIM and the pre-recorded resources reserved by other VNFs; and, if the reservable resources of a VNF exceed the resources required by the VNF, the NFVO reserving the required resources for the VNF and recording the correspondence between the VNF and the reserved resources.

[0005] For example, the invention patent with publication number CN120429116A discloses an adaptive cloud management platform system based on intelligent resource scheduling and container orchestration, which includes: a fine-grained resource scheduling and adaptive optimization module, a containerized application lifecycle management and dynamic container orchestration module, a high-precision operation and maintenance monitoring and self-healing mechanism module based on big data analysis, and a dynamic resource allocation and elastic scaling strategy module of the intelligent scheduling engine.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0007] In existing technologies, the demand for computing resources for a single simulation sample fluctuates dramatically at different simulation stages, and there will be periodic peaks in resource demand. When a large number of such samples are simulated in parallel, since their simulation progress is independent, their internal peak periods are randomly distributed in time and are very likely to overlap at a certain moment. This causes multiple samples to apply for peak resources from the resource pool at the same time, resulting in serious resource contention and low resource utilization. Summary of the Invention

[0008] This application provides a method and system for flexible resource orchestration of a distributed intelligent parallel simulation engine. This addresses the problem in existing technologies where the computational resource requirements of a single simulation sample fluctuate drastically at different simulation stages, resulting in periodic peaks in resource demand. When a large number of such samples are simulated in parallel, their internal peak periods are randomly distributed over time due to their independent simulation progress, potentially overlapping at a certain moment. This leads to multiple samples simultaneously requesting peak resources from the resource pool, causing severe resource contention and low resource utilization. By using dynamic load prediction and intelligent scheduling, the method effectively smooths out periodic peak resource demand in distributed simulations, significantly improving overall resource utilization.

[0009] This application provides a method for elastic resource orchestration of a distributed intelligent parallel simulation engine, including the following steps: performing dynamic load prediction based on simulation data generated by each simulation process in each simulation period in the distributed simulation simulation system, obtaining the predicted load state parameters of each simulation process in each subsequent simulation period, and dynamically marking the periods accordingly; dynamically setting scheduling instructions based on the dynamic marking results of the periods and simulation event rules, and sorting the dynamically set scheduling instructions based on the topological constraint characteristics of the distributed simulation simulation system, and executing optimized scheduling instructions based on the sorting results; and performing container status evaluation in real time based on the running status parameters of each simulation container during the simulation simulation, and performing dynamic fault handling based on the container status evaluation results and fault types.

[0010] This application provides an elastic resource orchestration system for a distributed intelligent parallel simulation engine, comprising a time period marking module, a resource scheduling module, and a fault handling module: The time period marking module is used to dynamically predict the load based on the simulation data generated by each simulation process within each simulation period in the distributed simulation system, obtaining the predicted load state parameters for each subsequent simulation period of each simulation process, and dynamically marking the time periods accordingly; The resource scheduling module is used to dynamically set scheduling instructions based on the time period dynamic marking results and simulation event rules, and to sort the dynamically set scheduling instructions based on the topological constraint characteristics of the distributed simulation system, executing optimized scheduling instructions based on the sorting results; The fault handling module is used to perform real-time container status evaluation based on the running status parameters of each simulation container during the simulation process, and to perform dynamic fault handling based on the container status evaluation results and fault types.

[0011] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0012] 1. This invention provides accurate and forward-looking decision-making basis for resource scheduling by integrating dynamic load prediction and time period marking of system, simulation and business data. Based on simulation event rules and topological constraint characteristics, it generates and optimizes scheduling instruction sequences, thereby achieving intelligent balancing and conflict avoidance of parallel simulation process resources, effectively smoothing peak resource demand. At the same time, through multi-layer fault real-time assessment and dynamic processing mechanisms, it ensures the continuity of system operation, thus achieving a significant improvement in the resource utilization of large-scale distributed simulation platforms and a comprehensive enhancement of the overall system reliability.

[0013] 2. By integrating a multi-mode adaptive prediction mechanism that combines periodic analysis, trend judgment, and machine learning prediction, and introducing a dynamic weight fusion strategy based on historical accuracy, the system can intelligently match the optimal prediction method, effectively cope with stable, periodic, and abrupt load fluctuations, thereby significantly improving prediction accuracy and robustness. Furthermore, by using sliding window and exponential smoothing techniques to avoid overfitting risks, the system ensures that the prediction model continuously adapts to real-time data characteristics, ultimately achieving a high-accuracy prediction of future load conditions and providing a reliable data foundation for resource scheduling decisions.

[0014] 3. By combining the dynamic marking results of time periods with simulation event rules, a preliminary scheduling strategy is generated. A multi-objective resource optimization algorithm is introduced to globally optimize the instruction set. The improved NSGA-III algorithm is used to generate the Pareto optimal solution set. Then, the TOPSIS method is used to select the best scheduling scheme that takes into account resource efficiency, cost, and performance, thereby achieving accurate quantification and intelligent selection of vertical / horizontal scaling instructions. Furthermore, the optimization scheme is transformed into specific resource orchestration actions based on the elastic scaling control algorithm, enabling the system to adaptively select the most suitable scaling strategy according to the load peak and valley characteristics. Ultimately, this significantly improves resource utilization efficiency and the scientific nature of scheduling decisions while ensuring the service quality of the simulation process.

[0015] 4. By integrating topological constraint features and obtaining a global resource view based on the Gossip protocol, scheduling instructions are finely classified and sorted in multiple dimensions, prioritizing the execution order of instructions with logical dependencies to avoid resource allocation conflicts. Simultaneously, by combining node link bandwidth and load status indices, differentiated sorting strategies are formulated for expansion and contraction instructions, effectively reducing node resource contention and network congestion risks. Finally, through a parallel and serial adaptive instruction execution mechanism, scheduling efficiency is improved while strictly controlling system concurrency pressure, thereby achieving the coordination, orderliness, and efficiency of resource scheduling in a large-scale distributed simulation environment, significantly improving the overall resource utilization efficiency and stability of the system. Attached Figure Description

[0016] Figure 1 Flowchart of the elastic resource orchestration method for the distributed intelligent parallel inference engine provided in this application embodiment;

[0017] Figure 2 The core flowchart of the elastic resource orchestration of the distributed intelligent parallel inference engine provided in the embodiments of this application is shown below.

[0018] Figure 3 This is a schematic diagram of the structure of the elastic resource orchestration system of the distributed intelligent parallel inference engine provided in the embodiments of this application. Detailed Implementation

[0019] This application provides a method and system for elastic resource orchestration of a distributed intelligent parallel inference engine. This addresses the problem in existing technologies where the computational resource requirements of a single inference sample fluctuate drastically at different simulation stages, resulting in periodic peaks in resource demand. Furthermore, when a large number of such samples are simulated in parallel, their internal peak periods are randomly distributed over time due to their independent simulation progress, potentially overlapping at certain moments. This leads to multiple samples simultaneously requesting peak resources from the resource pool, causing severe resource contention and low resource utilization. The overall approach is as follows:

[0020] By collecting and fusing multi-level data from the system, simulation, and business layers in real time, an adaptive prediction algorithm is used to accurately predict future loads. Based on this, subsequent simulation periods are dynamically marked as different levels, providing forward-looking data support for resource scheduling. Then, preliminary optimized scheduling instructions are dynamically generated according to simulation event rules. System topology constraints are introduced to finely sort and resolve conflicts among these instructions, forming an optimal instruction execution sequence. This ensures that scheduling actions meet both business needs and the objective constraints of the underlying architecture, achieving balanced and efficient resource allocation. During runtime, the health status of containers is continuously monitored. A multi-parameter fusion evaluation model is used to diagnose performance-level, container-level, node-level, and region-level anomalies in real time, automatically triggering corresponding graded fault self-healing processes. The entire process ensures state synchronization through a consistency protocol, achieving automated and zero-interruption fault handling. Ultimately, under large-scale, highly volatile loads, the optimal balance between resource utilization and business reliability is achieved.

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0022] like Figure 1 The diagram shown is a flowchart of the elastic resource orchestration method for a distributed intelligent parallel inference engine provided in this application embodiment. The method includes the following steps:

[0023] Dynamic load prediction is performed based on simulation data generated by each simulation process and simulation period in the distributed simulation system. This yields predicted load status parameters for each subsequent simulation period of each process, and time periods are dynamically marked accordingly, providing a basis for resource scheduling decisions. The distributed simulation system is a large-scale simulation platform supporting parallel simulation of multiple processes. The simulation data includes load status parameters and business semantic parameters. Based on the observation perspective and monitoring target level of the simulation data, it can be divided into system-level data, simulation-level data, and business-level data. System-level data, based on the infrastructure resource health perspective, reflects the resource consumption status of the underlying computing nodes, including node CPU / GPU utilization (Central Processing Unit Utilization / Graphics Processing - Unit Utilization), memory usage, and network I / O. Input / Output and disk I / O are considered. Simulation-level data, based on the application process runtime performance perspective, reflects the internal execution efficiency and status of a single simulation process, including the stepping rate of a single simulation process (whether it keeps up with real-time), critical thread load, message queue depth, and entity interaction event frequency. Business-level data, based on domain semantics and business process perspective, reflects the business logic status and hotspots of the simulation itself, including simulation phases and activity intensity in battlefield hotspot areas. Dynamic scheduling instructions are dynamically set based on time-period dynamic marking results and simulation event rules to ensure balanced resource allocation for parallel simulation processes. These dynamically set scheduling instructions are sorted based on the topological constraints of the distributed simulation system, and optimized scheduling instructions are executed based on the sorting results. Topological constraints define the multidimensional relationships and operational status attributes of simulation nodes at the structural and behavioral levels. During the simulation process, container status is evaluated in real-time based on the operational status parameters of each simulation container, and dynamic fault handling is performed based on the container status evaluation results and fault types. Fault types include performance-level anomalies, container-level anomalies, node-level anomalies, and region-level anomalies. Dynamic fault handling means dynamically executing corresponding fault handling procedures to improve the reliability of elastic resource orchestration.

[0024] In this embodiment, simulation data can be directly collected by a monitoring agent deployed within the simulation container. The core of the distributed simulation system lies in achieving efficient management and elastic scaling of multi-task parallel simulations through containerization and orchestration automation technologies. In this system, the basic hardware unit is the simulation node, i.e., a physical server or virtual machine, providing computing resources such as CPU, GPU, and memory. Lightweight, isolated runtime environments, i.e., simulation containers, are created on the simulation nodes using container engines such as Docker. The simulation process is the main program of a single simulation task instance. The system utilizes containerization technology to provide a lighter and more granular isolation layer on top of the nodes. Each simulation container is allocated a defined resource quota, ensuring that the resource usage of the simulation process within the container does not compete with processes in other containers on the same node, thus guaranteeing environmental consistency and stability. Multiple simulation containers can run simultaneously on a single simulation node, and each simulation container runs a simulation process. The simulation engine, as the core computing library running inside the container, actually drives the logical execution of the simulation process; and a simulation process can be divided into multiple simulation periods in the time dimension. This invention significantly improves the resource utilization efficiency and system reliability of large-scale distributed simulation platforms by constructing a complete closed-loop management system from prediction and scheduling to fault tolerance. Through multi-level data fusion analysis, it achieves accurate load prediction and dynamic time-period marking, providing a forward-looking decision-making basis for resource scheduling. Instruction sequencing based on topology constraints ensures that scheduling actions conform to system architecture characteristics, avoiding resource conflicts, effectively balancing multi-process resource allocation, and smoothing periodic peak demands, thereby greatly improving overall resource utilization. Simultaneously, the introduction of a tiered fault handling mechanism enables rapid response to various problems ranging from performance anomalies to regional faults, dynamically executing recovery processes, greatly enhancing the system's resilience and robustness in complex simulation scenarios, ultimately ensuring the efficient and stable execution of large-scale parallel simulation tasks.

[0025] like Figure 2 The diagram shown is a core flowchart of the elastic resource orchestration of the distributed intelligent parallel inference engine provided in this application embodiment. Its core implementation logic involves fine-grained collection of the running status at the system level, simulation level, and business level, real-time monitoring of key indicators, and then using rules or pre-trained algorithm models to perform dynamic resource scheduling and orchestration execution. At the underlying level, containerization, resource pooling, and lightweight simulation nodes are required to provide support to ensure dynamic load balancing at the system level, simulation level, and business level.

[0026] Furthermore, the methods for obtaining the predicted load state parameters for each simulation process and each subsequent simulation period are as follows: S1: Aggregate the load state parameters of each simulation process's historical simulation periods and the real-time load state parameters according to the simulation period to obtain the time-series data records for each simulation process; S2: If the load state parameters in the time-series data records exhibit periodic characteristics within the simulation period, calculate the predicted values ​​of the corresponding load state parameters based on these periodic characteristics; S3: If the slope change values ​​of the load state parameters in the time-series data records do not reach the preset slope change threshold in any adjacent simulation periods within the simulation period, or if the slope change values ​​in adjacent simulation periods reach the preset slope change threshold but the number of occurrences does not reach the preset number of occurrences threshold... If the average value of the load state parameter in each simulation period within the simulation cycle is recorded as the predicted value of the load state parameter corresponding to the simulation process; S4: If the slope change value of the load state parameter in the time series data record reaches the preset slope change threshold and the number of times it reaches the preset number threshold within the simulation period, then the feature data of the load state parameter in each simulation period is extracted and input into the preloaded data prediction model to obtain the predicted value of the load state parameter corresponding to the simulation process; where the data prediction model is an algorithm or neural network structure used to quantitatively predict future load state parameters based on historical time series data and real-time features, including LSTM (Long Short-Term Memory Network), etc. S5: If the load state parameters in the time series data record simultaneously meet the periodicity condition S2 and the slope change condition S3 within the simulation period, the obtained periodicity prediction value and mean prediction value are weighted and fused with the preset periodicity prediction weight and mean prediction weight respectively to obtain the predicted value of the load state parameter corresponding to the simulation process. The periodicity prediction weight and mean prediction weight are dynamically adjusted based on the historical prediction accuracy of the periodicity prediction value and mean prediction value of the simulation process. The system continuously calculates the mean square error of the periodic prediction and mean prediction in a fixed simulation period and dynamically allocates weights according to whether the periodic error is lower than the threshold proportion of the mean error: the prediction method with smaller error increases the weight according to the threshold proportion of the mean error. At the same time, exponential smoothing or sliding window is used to limit the range of historical data to avoid overfitting. Finally, normalization processing is used to ensure that the weight sum is 1, so that the fusion strategy can adaptively bias towards the prediction value with better historical performance, thereby improving the prediction robustness under complex load scenarios.

[0027] Specifically, the steps for dynamic marking of time periods include: predicting load status parameters, including basic resource parameters and service performance parameters; calculating the load status evaluation index for each simulation process and each subsequent simulation period based on the service performance parameters of each simulation process and each subsequent simulation period; if the load status evaluation index of a simulation process and a subsequent simulation period is lower than a preset first load status threshold, or if its basic resource parameters are lower than the corresponding first basic resource parameter threshold, then the simulation period is marked as the first time period; if the load status evaluation index of a simulation process and a subsequent simulation period is not lower than the preset first load status threshold, but does not exceed the preset second load status threshold, or if its basic resource parameters are not lower than the corresponding first basic resource parameter threshold and do not exceed the corresponding second basic resource parameter threshold, then the simulation period is marked as the second time period; if the load status evaluation index of a simulation process and a subsequent simulation period exceeds the preset second load status threshold, or if its basic resource parameters exceed the corresponding second basic resource parameter threshold, then the simulation period is marked as the third time period.

[0028] The basic resource parameters include CPU utilization, GPU utilization, memory usage, message queue depth, and network I / O throughput. Service performance parameters include step rate, entity interaction event frequency, and activity intensity in each region. Step rate is defined as the ratio of the rate of time progression in the simulation world to the rate of time elapsed in the real world. Ideally, the step rate should be 1, indicating that the simulation timing is synchronized with the real-time clock. If the ratio is less than 1, the simulation calculation lags behind the real-time clock; if the ratio is greater than 1, the simulation calculation is ahead of the real-time clock. Entity interaction event frequency refers to the number of effective interactions between simulated entities per unit time. Typical types include sensor detection, missile launch, communication interaction, collision detection, and damage calculation. This frequency is the main driving factor for the system's computational load. Activity intensity in each region is a quantitative indicator of the density of entity activities and events in different geographical areas within the simulation world. It is calculated by dividing the simulation battlefield into grid cells of fixed area. The sum of the products of the weight coefficients of all events in each cell and their corresponding occurrence counts represents the activity intensity of that region. The setting of the event weight coefficients depends on the event type and its consumption of computational resources. The load status evaluation index for each simulation process and each subsequent simulation period is obtained as follows: The step rate reference value, the entity interaction event frequency for each simulation process and each subsequent simulation period, and the activity intensity for each region are each compared to their corresponding step rate, entity interaction event frequency reference value, and activity intensity reference value. Then, the results of these comparisons are weighted using the business performance parameter weighting factors. Finally, the weighted results are coupled together (summed) to obtain the load status evaluation index for each simulation process and each subsequent simulation period. The business performance parameter weighting factors include the step rate weighting factor, the entity interaction event frequency weighting factor, and the activity intensity weighting factor; the load status evaluation index is a comprehensive quantitative numerical indicator of the load status based on the integrated processing of business performance parameters.

[0029] In this embodiment, dynamic load prediction is based on a simulation cycle. The simulation cycle is a preset, repeatable time interval, the length of which is determined by business needs or system characteristics, such as the daily traffic cycle of e-commerce or the daily operation cycle of industrial equipment. A single simulation cycle can be divided into multiple simulation periods, each representing a slice of the system state used to capture dynamic load changes. The load state parameters of historical simulation periods, real-time load state parameters, and load state parameters of each subsequent simulation period are all within the same simulation cycle of the same simulation process. This invention predicts the load data for the remaining periods within the same cycle based on the load data (historical load data and real-time load data) of the simulation periods that have already occurred within the simulation cycle, ensuring that the prediction model can capture recurring patterns within the cycle. At the end of each cycle, complete cycle data can be fed back to the model to optimize the prediction accuracy for the next cycle. This invention accurately obtains future load status through an adaptive prediction strategy that integrates multi-source time-series data, and constructs a comprehensive evaluation system based on multi-dimensional business and resource indicators to intelligently mark time periods, thereby providing a high-precision and forward-looking decision-making basis for resource scheduling. Furthermore, by combining dynamic weight adjustment and fault hierarchical processing mechanisms, the system can adaptively optimize prediction accuracy and quickly respond to various anomalies, ultimately achieving efficient smoothing of resource demand and a significant improvement in system reliability under complex simulation environments.

[0030] Furthermore, the steps for dynamically setting scheduling instructions based on the time-period dynamic marking results and simulation event rules include: scheduling instructions include vertical shrinkage instructions, horizontal shrinkage instructions, vertical expansion instructions, and horizontal expansion instructions; the scheduling instruction set is processed based on a multi-objective resource optimization algorithm to obtain an optimized scheduling instruction set. This process uses an improved NSGA-III (Non-dominated Sorting Genetic Algorithm III) multi-objective evolutionary algorithm to process the initial scheduling instruction set. This algorithm generates a Pareto optimal solution set through efficient exploration of a large-scale scheduling scheme space. Each solution in this set represents the optimal trade-off among multiple competing objectives, and the objective function is defined as minimizing the weighted comprehensive index: ;

[0031] in, For resource costs, For end-to-end task latency, As an energy consumption indicator, , and These are the corresponding dynamic weight coefficients. The algorithm must satisfy the following: the completion time of all tasks or workloads must not be later than their preset latest completion time limit; the total allocation of any resource type (such as CPU, memory) among all instances or tasks must not exceed the total available capacity of that resource in the cloud platform or cluster; the system must ensure that the overall service availability is not less than 99.9%. Subsequently, based on this Pareto solution set, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method is used for the final scheduling decision. By calculating the relative proximity between each non-dominated solution and the ideal solution and the negative ideal solution, the solution that is closest to the ideal solution and furthest from the negative ideal solution is selected as the optimal scheduling strategy. To further improve the stability and response quality of resource adjustment, the system also introduces an elastic scaling strategy based on the proportional-integral-derivative (PID) control principle: Let the current resource utilization rate u(t) be , and the target utilization rate be u target =70%; the control error is: e(t) = u(t) – u target The resource adjustment amount based on PID control is: ;in, , and These are proportional gain, integral gain, and derivative gain, respectively. By adjusting the PID parameters and combining them with the load change rate to dynamically adjust the gain coefficient, the resource utilization rate can be stabilized within a limited range of the target value, reducing ineffective scaling operations. Finally, the system combines the selected TOPSIS optimal solution with the PID control strategy to output a set of scheduling instructions that comprehensively consider multi-objective optimization and dynamic response quality, namely, the optimized scheduling instruction set. Based on the simulation events in each subsequent simulation period of each simulation process, the corresponding optimized scheduling instructions are set according to the preset simulation event rules. If the number of simulation periods that are third periods in the subsequent simulation periods of a certain simulation process exceeds the preset threshold for the number of periods, then optimized vertical expansion instructions are set for each subsequent simulation period of that simulation process. If there are third periods in the subsequent simulation periods of a certain simulation process, and the number of simulation periods that are third periods does not exceed the preset threshold for the number of periods, then optimized horizontal expansion instructions are set for each subsequent simulation period of that simulation process. If there are no third periods in the subsequent simulation periods of a certain simulation process, but there are first periods and not all simulation periods are first periods, then optimized vertical shrinkage instructions are set for each subsequent simulation period of that simulation process. If all simulation periods in the subsequent simulation periods of a certain simulation process are first periods, then optimized horizontal shrinkage instructions are set for each subsequent simulation period of that simulation process.

[0032] In this embodiment, the resource elastic scheduling strategy in the distributed simulation and simulation system is mainly divided into two categories: vertical scaling and horizontal scaling. Vertical scaling enhances the processing capacity of a single simulation process instance by increasing the physical resources it occupies, such as CPU, memory, and disk I / O, and is suitable for handling continuous and stable high-load scenarios. After receiving the instruction, the system will restart the instance according to the predetermined new quota to apply the configuration. Horizontal scaling distributes simulation tasks to multiple identical instances for parallel processing through a load balancer, thereby improving the overall system throughput and enabling rapid response to sudden traffic surges. It is suitable for handling short-term or sudden high loads. Newly launched simulation and simulation instances are pulled and deployed by the system from the image repository and automatically registered to the service pool. The scaling strategy is the same: vertical scaling reduces the physical resources occupied by a single simulation process instance while retaining the instance architecture to maintain system elasticity, and is suitable for scenarios where the overall load is reduced but still fluctuates; horizontal scaling directly reduces the number of instances and is the most effective cost optimization method for handling continuous and stable low loads. The system achieves this by removing instances and releasing resources, striking a balance between cost savings and maintaining elasticity.

[0033] Furthermore, the steps for sorting dynamically set scheduling instructions based on the topology constraint features of the distributed simulation system include: topology constraint features include physical connections, logical dependencies, and load states between simulation nodes; classifying the scheduled instructions to be executed for each subsequent simulation period to obtain a set of scheduled instructions to be executed for each subsequent simulation period; obtaining the physical connection state and load state of each simulation node based on the global resource view obtained through the Gossip protocol; for each simulation node in the set of scheduled instructions to be executed for the same subsequent simulation period that has logical dependencies, sorting the corresponding instructions based on logical dependency rules, for example, if the execution of instruction A is a prerequisite for the effectiveness of instruction B, then instruction A is placed before instruction B, and logical dependencies create a fundamental constraint on the order of instructions; and mapping each simulation node to each simulation process. The load status evaluation index of each simulation node in each subsequent simulation period is obtained by determining the relationship between the two nodes. For the optimized expansion instructions in the set of optimized scheduling instructions to be executed in the same subsequent simulation period, for simulation nodes with physical connections, the corresponding instructions are sorted in ascending order based on the bandwidth utilization of the links between nodes, and then all optimized expansion instructions are sorted in ascending order based on the load status evaluation index to obtain the first instruction sorting sequence. For the optimized shrinkage instructions in the set of optimized scheduling instructions to be executed in the same subsequent simulation period, all optimized shrinkage instructions are sorted in descending order based on the load status evaluation index to obtain the second instruction sorting sequence. The first instruction sorting sequence and the second instruction sorting sequence corresponding to the same subsequent simulation period are merged and sorted in ascending order based on the load status evaluation index to obtain the final instruction execution sequence corresponding to each subsequent simulation period.

[0034] The steps for executing optimized scheduling instructions based on the sorting results include: if the number of optimized scheduling instructions in the same subsequent simulation period does not exceed the preset parallel instruction threshold, then each optimized scheduling instruction is executed in parallel; if the number of optimized scheduling instructions in the same subsequent simulation period exceeds the preset parallel instruction threshold, then each optimized scheduling instruction is executed sequentially based on the corresponding final instruction execution sequence.

[0035] In this embodiment, after generating the instruction sequence, its feasibility needs to be verified based on the global resource view: if it is determined to be feasible, the sequence is output as the final result and executed; if it is determined to be infeasible due to insufficient target node resources, network link failure, or other reasons, an adjustment mechanism is triggered. The adjustment strategies include: instruction replacement, i.e., selecting a candidate instruction with lower priority but feasible to replace the original infeasible instruction; node replacement, i.e., finding other alternative nodes that meet the topological constraints for the same instruction, recalculating their priority, and inserting them into the sequence; sequence rearrangement, i.e., skipping the current infeasible instruction and executing subsequent instructions, while moving the current instruction to the waiting queue for re-evaluation in the next scheduling cycle; and forced execution, i.e., if it is a critical instruction, triggering the underlying resource expansion operation before execution. This invention integrates topological constraint features and a global resource view obtained based on the Gossip protocol to perform refined classification and multi-dimensional sorting of scheduling instructions, prioritizing the execution order of instructions with logical dependencies to avoid resource allocation conflicts. Simultaneously, it combines node link bandwidth and load status indices to formulate differentiated sorting strategies for expansion and contraction instructions, effectively reducing node resource contention and network congestion risks. Finally, through a parallel and serial adaptive instruction execution mechanism, it improves scheduling efficiency while strictly controlling system concurrency pressure, thereby achieving a collaborative, orderly, and efficient resource scheduling process in a large-scale distributed simulation environment, significantly improving the overall resource utilization efficiency and stability of the system.

[0036] Furthermore, the operational status parameters include average response time, service response success rate, number of check failures, and the historical average load status assessment index of the corresponding simulation process. The average response time is the average time for the simulation container to process each simulation event historically. The service response success rate is the ratio of the number of simulation requests successfully processed by the simulation container to the total number of requests within a historical simulation period. The number of check failures is the total number of liveness probe and ready probe check failures of the simulation container within a historical simulation period, obtained directly from the system logs. The steps for evaluating the container status based on the operational status parameters of each simulation container include: calculating the ratios of the average response time reference value, the service response success rate reference value, and the number of check failures reference value of each simulation container to their corresponding average response time, service response success rate reference value, and number of check failures reference value, respectively, to obtain the response time impact parameter, service response success rate impact parameter, and check failures impact parameter of each simulation container; and comparing the allowable deviation average load status assessment index with the historical average load status index of each simulation container. The deviation compliance of the load status assessment index and the average load status assessment index reference values ​​is calculated to obtain the average load status impact parameter for each simulated container. The response time impact parameter, service response success rate impact parameter, check failure count impact parameter, and average load status impact parameter for each simulated container are weighted using operational status parameter weighting factors. The weighting results are then coupled to obtain the operational status assessment index for each simulated container. If the operational status assessment index of a simulated container does not reach the preset operational status assessment threshold, the simulated container is marked as an abnormal container; if the operational status assessment index of a simulated container reaches the preset operational status assessment threshold, the simulated container is marked as a normal container. The operational status parameter weighting factors include response time weighting factor, service response success rate weighting factor, check failure count weighting factor, and historical average load weighting factor. The operational status assessment index is a comprehensive quantitative numerical indicator of the container's health and performance status based on the integrated processing of operational status parameters.

[0037] The steps for dynamic fault handling based on container status assessment results and fault types include: marking simulation containers whose failure count exceeds a preset failure count threshold as container-level abnormal containers; marking simulation containers whose associated simulation nodes malfunction as node-level abnormal containers; marking simulation containers whose data centers or availability zones experience failures as region-level abnormal containers; marking simulation containers that do not meet the container-level, node-level, and region-level abnormal judgment criteria, but whose operational status assessment index does not reach a preset operational status assessment threshold, as performance-level abnormal containers; if a simulation container is a container-level abnormal container, the container-level abnormal handling process is initiated: restarting the simulation container until its operational status assessment index reaches the operational status assessment threshold; if the number of restarts reaches a preset restart count threshold, a container abnormality alarm is generated and reported to designated personnel; if a simulation container is a node-level abnormal container, the node-level abnormal handling process is initiated: the simulation container is restarted... All tasks on the simulation node where the real container resides are migrated to healthy nodes within the backup cluster. This process employs a macro- and micro-level collaborative scheduling strategy: at the macro level, computing resources are globally allocated through a cross-resource pool load balancing algorithm, such as dynamically expanding the electronic warfare resource pool by two replicas; at the micro level, fine-grained resource quota adjustments are made for individual containers, such as limiting the CPU usage of each replica to no more than two cores, thereby optimizing resource utilization while ensuring business continuity; if the simulation container is a region-level abnormal container, a region-level abnormality handling process is initiated: the traffic scheduling system switches all business traffic to the off-site disaster recovery center within a recovery time target of less than 30 seconds; if the simulation container is a performance-level abnormal container, a performance-level abnormality handling process is initiated: the status and response time of the external services that the container depends on are automatically detected, and if an abnormality is identified in the dependent service, an alarm message for that service is generated and issued.

[0038] In this embodiment, the present invention constructs a comprehensive evaluation system of multi-dimensional operational status parameters, enabling the system to accurately quantify container health status and intelligently identify anomalies. Based on a hierarchical fault classification mechanism including performance level, container level, node level, and region level, and a dynamic processing flow, it achieves refined fault handling from container restart and resource scheduling to cross-regional traffic switching. During dynamic fault handling, the Raft algorithm is used to ensure state consistency during fault switching, ensuring zero task migration loss. Combining macro- and micro-level collaborative scheduling strategies, it achieves both cross-resource pool load balancing and resource quota optimization for individual containers. Ultimately, a complete fault-tolerant system integrating early anomaly detection, intelligent fault location, hierarchical self-healing processing, and high availability assurance is formed, significantly improving the system's robustness, availability, and business continuity in complex simulation environments.

[0039] like Figure 3The diagram shows the structure of the elastic resource orchestration system of the distributed intelligent parallel simulation engine provided in this application embodiment, including a time period marking module, a resource scheduling module, and a fault handling module: The time period marking module is used to dynamically predict the load based on the simulation data generated by each simulation process within each simulation period in the distributed simulation simulation system, obtain the predicted load state parameters for each subsequent simulation period of each simulation process, and dynamically mark the time periods accordingly; The resource scheduling module is used to dynamically set scheduling instructions based on the time period dynamic marking results and simulation event rules, and sort the dynamically set scheduling instructions based on the topological constraint characteristics of the distributed simulation simulation system, and execute optimized scheduling instructions based on the sorting results; The fault handling module is used to perform real-time container status evaluation based on the running status parameters of each simulation container during the simulation simulation process, and perform dynamic fault handling based on the container status evaluation results and fault types.

[0040] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0041] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0042] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0043] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0044] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0045] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for elastic resource orchestration in a distributed intelligent parallel inference engine, characterized in that, Includes the following steps: Dynamic load prediction is performed based on the simulation data generated by each simulation process and simulation period in the distributed simulation system. The predicted load state parameters of each simulation process and each subsequent simulation period are obtained, and the period is dynamically marked accordingly. The distributed simulation system is a large-scale simulation platform that supports parallel simulation of multiple simulation processes. The simulation data includes load state parameters and business semantic parameters. The steps for performing dynamic time-period marking include: The predicted load status parameters include basic resource parameters and service performance parameters. The basic resource parameters include CPU utilization, GPU utilization, memory usage, message queue depth, and network I / O throughput. The service performance parameters include step rate, entity interaction event frequency, and activity intensity in each region. The load status evaluation index for each simulation process and each subsequent simulation period is calculated based on the business performance parameters of each simulation process and each subsequent simulation period. If the load status evaluation index of a certain simulation process in a subsequent simulation period is lower than the preset first load status threshold, or if there are basic resource parameters that are lower than the corresponding first threshold of basic resource parameters, then the simulation period is marked as the first period. If the load status evaluation index of a certain simulation process in a subsequent simulation period is not lower than the preset first load status threshold, but does not exceed the preset second load status threshold, or if its basic resource parameters are not lower than the corresponding first basic resource parameter threshold and do not exceed the corresponding second basic resource parameter threshold, then the simulation period is marked as the second period. If the load status evaluation index of a certain simulation process in a subsequent simulation period exceeds the preset second threshold of load status, or if its basic resource parameters exceed the corresponding second threshold of basic resource parameters, then the simulation period is marked as the third period. The optimized scheduling instructions are dynamically set based on the time period dynamic marking results and simulation event rules, and the dynamically set scheduling instructions are sorted based on the topological constraint features of the distributed simulation simulation system. The optimized scheduling instructions are executed based on the sorting results. The topological constraint features define the multidimensional correlation and running status attributes of simulation nodes at the structural and behavioral levels. During the simulation, the container status is evaluated in real time based on the running status parameters of each simulation container, and dynamic fault handling is performed based on the container status evaluation results and fault types. The fault types include performance-level anomalies, container-level anomalies, node-level anomalies, and region-level anomalies. The dynamic fault handling means dynamically executing the corresponding fault handling process to improve the reliability of elastic resource orchestration.

2. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 1, characterized in that, The predicted load state parameters for each subsequent simulation period in each simulation process are obtained as follows: S1: Aggregate the load status parameters of each simulation process during historical simulation periods and the real-time load status parameters according to the simulation period to obtain the time series data records of each simulation process; S2: If the load state parameters in the time series data record have periodic characteristics within the simulation period, then calculate the predicted value of the corresponding load state parameters based on the periodic characteristics. S3: If the slope change value of the load state parameter in the time series data record does not reach the preset slope change threshold in each adjacent simulation period within the simulation cycle, or if the slope change value of adjacent simulation periods reaches the preset slope change threshold but the number of times does not reach the preset number of times threshold, then the average value of the load state parameter in each simulation period within the simulation cycle is recorded as the predicted value of the load state parameter corresponding to the simulation process. S4: If the slope change value of the load state parameter in the time series data record reaches a preset slope change threshold and the number of times it reaches a preset number threshold in adjacent simulation periods within the simulation cycle, then the feature data of the load state parameter in each simulation period is extracted and input into the preloaded data prediction model to obtain the predicted value of the load state parameter corresponding to the simulation process. S5: If the load state parameters in the time series data record simultaneously meet the periodicity feature condition described in S2 and the slope change condition described in S3 within the simulation period, the obtained periodicity feature prediction value and mean prediction value are weighted and fused with the preset periodicity feature weight and mean prediction weight respectively to obtain the predicted value of the load state parameters corresponding to the simulation process.

3. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 1, characterized in that, The load status evaluation index for each subsequent simulation period in each simulation process is obtained as follows: The step rate reference value, the entity interaction event frequency of each simulation process and the activity intensity of each region are respectively compared with the corresponding step rate, entity interaction event frequency reference value and activity intensity reference value. Then, the results of the comparison value calculation using the business efficiency parameter weight factor are weighted and processed. Finally, the weighting results are coupled to obtain the load status evaluation index of each simulation process and each subsequent simulation period. The business performance parameter weighting factors include step rate weighting factor, entity interaction event frequency weighting factor, and activity intensity weighting factor. The load status assessment index is a numerical indicator of the comprehensive quantitative load status based on the integrated processing of business performance parameters.

4. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 1, characterized in that, The steps for dynamically setting scheduling instructions based on time-period dynamic marking results and simulation event rules include: The scheduling instructions include vertical shrinkage instructions, horizontal shrinkage instructions, vertical expansion instructions, and horizontal expansion instructions; An optimized scheduling instruction set is obtained by processing the scheduling instruction set based on a multi-objective resource optimization algorithm; Based on the simulation events in each subsequent simulation period of each simulation process, corresponding optimized scheduling instructions are set according to the preset simulation event rules. If the number of simulation periods that are the third period in the subsequent simulation periods of a certain simulation process exceeds the preset threshold for the number of periods, then an optimized vertical expansion instruction is set for each subsequent simulation period of that simulation process. If a third time period exists in the subsequent simulation time period of a certain simulation process, and the number of simulation time periods that are third time periods does not exceed the preset time period number threshold, then an optimized horizontal expansion instruction is set for each subsequent simulation time period of the simulation process. If there is no third time period in the subsequent simulation time periods of a certain simulation process, but there is a first time period and the simulation time periods are not all the first time period, then set the optimized vertical scaling instruction for each subsequent simulation time period of the simulation process. If all subsequent simulation periods of a certain simulation process are the first simulation period, then an optimized horizontal scaling instruction is set for each subsequent simulation period of that simulation process.

5. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 4, characterized in that, The steps for sorting dynamically set scheduling instructions based on the topology constraint features of the distributed simulation and deduction system include: The topology constraint features include the physical connections, logical dependencies, and load states between simulation nodes; Based on the subsequent simulation periods, the optimized scheduling instructions to be executed are classified to obtain the set of optimized scheduling instructions to be executed for each subsequent simulation period; For each simulation node with logical dependencies in the set of optimized scheduling instructions to be executed in the same subsequent simulation period, the corresponding instructions are sorted based on the logical dependency rules. Based on the mapping relationship between each simulation node and each simulation process, the load status evaluation index of each simulation node in each subsequent simulation period is obtained. For the optimized expansion type instructions in the set of optimized scheduling instructions to be executed in the same subsequent simulation period, for simulation nodes with physical connection relationships, the corresponding instructions are sorted in ascending order based on the bandwidth utilization of the link between nodes, and then all optimized expansion type instructions are sorted in ascending order based on the load status evaluation index to obtain the first instruction sorting sequence. For the optimized shrinkage type instructions in the set of optimized scheduling instructions to be executed in the same subsequent simulation period, all optimized shrinkage type instructions are sorted in descending order based on the load status evaluation index to obtain the second instruction sorting sequence. After merging the first instruction sorting sequence and the second instruction sorting sequence corresponding to the same subsequent simulation period, sort them in ascending order based on the load state evaluation index to obtain the final instruction execution sequence corresponding to each subsequent simulation period.

6. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 5, characterized in that, The step of executing optimized scheduling instructions based on the sorting results includes: If the number of optimized scheduling instructions in the same subsequent simulation period does not exceed the preset parallel instruction threshold, then each optimized scheduling instruction will be executed in parallel. If the number of optimized scheduling instructions in the same subsequent simulation period exceeds the preset parallel instruction threshold, then each optimized scheduling instruction will be executed sequentially based on the corresponding final instruction execution sequence.

7. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 1, characterized in that, The operational status parameters include average response time, service response success rate, number of failed checks, and the historical average load status evaluation index of the corresponding simulation process. The steps for evaluating the container status based on the operating status parameters of each simulation container include: The average response time reference value, the service response success rate reference value and the number of check failures reference values ​​of each simulation container are compared with their corresponding average response time, service response success rate reference value and number of check failures reference values ​​to obtain the response time influence parameter, service response success rate influence parameter and check failure number influence parameter of each simulation container. The deviation compliance of the allowable deviation average load state evaluation index with the historical average load state evaluation index and the reference value of the average load state evaluation index of each simulation container is calculated to obtain the average load state influence parameter of each simulation container. The impact parameters of response time, service response success rate, number of failed checks, and average load status of each simulation container are weighted using the weighting factors of the running status parameters. The weighting results are then coupled to obtain the running status evaluation index of each simulation container. If the simulation container's operational status evaluation index does not reach the preset operational status evaluation threshold, then the simulation container will be marked as an abnormal container. If the operating status evaluation index of the simulation container reaches the preset operating status evaluation threshold, the simulation container will be marked as a normal container. The weighting factors for the operational status parameters include response time weighting factor, service response success rate weighting factor, number of failed checks weighting factor, and historical average load weighting factor. The operational status assessment index is a numerical indicator that comprehensively quantifies the health and performance status of a container based on the integrated processing of operational status parameters.

8. The elastic resource orchestration method for the distributed intelligent parallel inference engine as described in claim 7, characterized in that, The steps for dynamic fault handling based on container status assessment results and fault types include: The simulation container whose number of failures exceeds a preset failure threshold will be marked as a container-level abnormal container. The simulation container whose associated simulation node is abnormal is marked as a node-level abnormal container; Mark any simulation container that fails in its data center or availability zone as a region-level anomalous container; Simulation containers that do not meet the criteria for container-level, node-level, and region-level anomalies, but whose operational status evaluation index does not reach the preset operational status evaluation threshold, are marked as performance-level anomaly containers. If the simulation container is a container-level exception container, then the container-level exception handling process is initiated. If the simulation container is a node-level exception container, then the node-level exception handling process is initiated. If the simulation container is a region-level exception container, then the region-level exception handling process is initiated. If the simulation container is a performance-level exception container, then the performance-level exception handling process will be initiated.

9. A flexible resource orchestration system for a distributed intelligent parallel inference engine, characterized in that, It includes a time period marking module, a resource scheduling module, and a fault handling module: The time period marking module is used to perform dynamic load prediction based on the simulation data generated by each simulation process in each simulation period in the distributed simulation system, to obtain the predicted load state parameters of each simulation process in each subsequent simulation period, and to perform dynamic time period marking accordingly. The steps for performing dynamic time period marking include: The predicted load status parameters include basic resource parameters and service performance parameters. The basic resource parameters include CPU utilization, GPU utilization, memory usage, message queue depth, and network I / O throughput. The service performance parameters include step rate, entity interaction event frequency, and activity intensity in each region. The load status evaluation index for each simulation process and each subsequent simulation period is calculated based on the business performance parameters of each simulation process and each subsequent simulation period. If the load status evaluation index of a certain simulation process in a subsequent simulation period is lower than the preset first load status threshold, or if there are basic resource parameters that are lower than the corresponding first threshold of basic resource parameters, then the simulation period is marked as the first period. If the load status evaluation index of a certain simulation process in a subsequent simulation period is not lower than the preset first load status threshold, but does not exceed the preset second load status threshold, or if its basic resource parameters are not lower than the corresponding first basic resource parameter threshold and do not exceed the corresponding second basic resource parameter threshold, then the simulation period is marked as the second period. If the load status evaluation index of a certain simulation process in a subsequent simulation period exceeds the preset second threshold of load status, or if its basic resource parameters exceed the corresponding second threshold of basic resource parameters, then the simulation period is marked as the third period. The resource scheduling module is used to dynamically set scheduling instructions based on the time period dynamic marking results and simulation event rules, sort the dynamically set scheduling instructions based on the topological constraint characteristics of the distributed simulation and deduction system, and execute optimized scheduling instructions based on the sorting results. The fault handling module is used to evaluate the container status in real time based on the operating status parameters of each simulation container during the simulation process, and to perform dynamic fault handling based on the container status evaluation results and fault types.

Images