Computing power network resource reconstruction method and system based on multi-element fusion, and storage medium

Abstract

The invention discloses a computing power network resource reconstruction method and system based on multi-element fusion and a storage medium, and the method comprises the steps: generating a resource list containing four elements of computing power, an algorithm, data and a model based on idle resources in a computing power network; analyzing a coupling relationship among elements in the resource list, and generating a resource coupling relationship graph; obtaining a task prediction list; generating a resource reconstruction strategy according to the resource coupling relation graph and the task prediction list; and in response to the state change in the computing power network, triggering and executing a resource reconstruction operation.

Description

Technical Field

[0001] This invention relates to the field of computing power network resource management technology, specifically to a method, system, and storage medium for reconstructing computing power network resources based on multi-element fusion. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence technologies, computing power networks have attracted widespread attention as a new type of network architecture. Computing power networks aim to deeply integrate multi-dimensional resources such as data, algorithms, computing power, and models to achieve dynamic collaboration and efficient utilization of resources. However, the resources in computing power networks exhibit significant heterogeneity and complexity. For example, computing power resources include general-purpose computing power, intelligent computing power, and supercomputing power; data resources include structured and unstructured data; and algorithms and models are extremely diverse.

[0003] Currently, the resource management methods for computing power networks mainly suffer from the following problems:

[0004] (1) Existing methods mostly reconstruct single elements (such as optimizing data storage or adjusting computing power allocation), without fully considering the inherent coupling relationship between computing power, algorithms, data and models, resulting in low overall resource utilization and difficulty in adapting to complex and ever-changing comprehensive task requirements;

[0005] (2) Resource matching strategies are mostly static or reactive, which cannot predict future needs based on historical task patterns and perform forward-looking resource optimization, making it difficult to adapt to the dynamic, diverse, and large-scale business scenarios of computing power networks.

[0006] (3) The lack of a real-time and continuous monitoring mechanism for resource and task status makes it impossible to quickly trigger the reorganization of resources when new resources are registered or new tasks arrive, resulting in insufficient system adaptability.

[0007] Therefore, there is an urgent need for a computing network resource management method that can integrate multiple factors and support dynamic prediction and real-time reconstruction in order to improve the flexibility, adaptability and overall efficiency of resource utilization. Summary of the Invention

[0008] This invention aims to provide a computing network resource reconstruction scheme based on multi-element fusion, in order to solve the problems of fragmented resource elements, lack of dynamic prediction and real-time reconstruction capabilities in the existing technology, and improve the adaptability and resource utilization efficiency of computing networks in the face of complex heterogeneous tasks.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for reconstructing computing network resources based on multi-element fusion, comprising:

[0011] Based on the idle resources in the computing power network, a resource list containing four categories of elements is generated;

[0012] Analyze the coupling relationships between the elements in the resource list to generate a resource coupling relationship map; obtain a task prediction list;

[0013] Based on the resource coupling relationship graph and the task prediction list, a resource reconfiguration strategy is generated;

[0014] In response to state changes in the computing network, resource reconfiguration operations are triggered and executed.

[0015] Secondly, the present invention provides a computing power network resource reconstruction system based on multi-element fusion, comprising:

[0016] The resource inventory management unit is used to generate and maintain a resource inventory that includes four categories of elements: computing power, algorithms, data, and models.

[0017] The coupling analysis unit is used to analyze the coupling relationships between resource elements and generate a resource coupling relationship map;

[0018] The task prediction unit is used to generate a task prediction list.

[0019] The strategy generation unit is used to generate a resource reconstruction strategy based on the graph and the prediction list;

[0020] The monitoring and execution unit is used to trigger and control the resource reconfiguration process in response to changes in the state of resources or tasks.

[0021] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0022] The beneficial effects that the multi-element fusion-based computing network resource reconstruction method, system, and storage medium disclosed in this application may bring include, but are not limited to:

[0023] 1. It has solved the problem of fragmented resource elements and achieved deep collaboration and efficient utilization of multi-dimensional resources.

[0024] Existing technologies often focus on independent optimization (single-element reconstruction) of single elements such as computing power, algorithms, data, or models, or only reconstruct at the hardware circuit level, neglecting the inherent coupling relationships between different elements in the computing power network. This leads to low overall resource utilization and difficulty in handling complex tasks requiring comprehensive optimization. This invention, by constructing a resource coupling relationship graph, is the first to systematically and quantitatively analyze the compatibility, execution efficiency, and synergistic effects between core elements such as data and algorithms, computing power and algorithms, and algorithms and models, transforming discrete resources into a network reflecting their inherent connections. The resource reconstruction strategy generated based on this graph can guide the system to intelligently combine highly coupled heterogeneous resources (such as specific algorithms with suitable computing power and matching datasets) into functional units, thereby breaking down barriers between elements and achieving a synergistic effect of "1+1>2." This significantly improves the efficiency and success rate of complex task execution, fundamentally overcoming the shortcomings of traditional methods in terms of low resource adaptability.

[0025] 2. It has achieved a shift from passive response to proactive adaptation, significantly improving the forward-looking nature and task matching flexibility of the computing network.

[0026] Traditional resource matching methods typically only begin searching for available resources after a task arrives, representing a passive and static response mode that cannot predict or adapt to dynamic, diverse, and large-scale task flows. This invention innovatively introduces a machine learning-based task prediction mechanism. By analyzing historical task information, it can predict task demands (including type, scale, and resource requirements) over a future period and generate a task prediction list. This allows the system to proactively generate or optimize resource reconfiguration strategies based on resource coupling relationship graphs before tasks actually arrive, enabling pre-configuration or pre-optimization of resources. When predicted peak tasks or new types of tasks actually arrive, the system is already "prepared," able to quickly provide highly suitable resource combinations, significantly reducing task queuing time and initialization overhead, smoothing resource load, and thus significantly enhancing the flexibility and agility of the computing network in responding to dynamic and diverse business scenarios.

[0027] 3. A dynamic closed loop of "perception-decision-execution" has been constructed, which greatly enhances the system's real-time responsiveness and resource utilization flexibility.

[0028] The environment of a computing network (resource status, task load) is constantly changing, while existing reconstruction strategies are often static or slowly updated. This invention establishes a complete adaptive closed loop through real-time monitoring and dynamic triggering mechanisms. On one hand, the system continuously monitors resource registration. When new resources reach a threshold, it automatically triggers updates to the resource list, resource coupling relationship graph, and even resource reconstruction strategies, ensuring that the system's understanding is always synchronized with the latest resource status. On the other hand, the system monitors the arrival of new tasks in real time and immediately assesses their coupling degree with the current resource structure. Once it falls below a performance threshold, it automatically triggers the reconstruction process, invoking the latest strategy for resource reorganization. This ability to "continuously perceive state changes, evaluate matching degree in real time, and dynamically trigger optimized reconstruction" enables the computing network to adjust its resource form in real time as the internal and external environment changes, like an organism, maintaining a high degree of adaptability and resource utilization flexibility, effectively solving the problem of insufficient adaptability of traditional methods.

[0029] 4. Improved overall resource utilization and the quality of service (QoS) of the computing network.

[0030] In summary, this invention, through deep collaboration, forward-looking prediction, and dynamic reconfiguration, enables the fuller and more rational utilization of limited, heterogeneous physical resources. Resources are no longer isolated and rigid, but can be flexibly decomposed and combined into various virtual functional units according to task requirements. This reduces reliance on expensive, single-purpose resources and improves the reuse rate of existing resources and overall output. Simultaneously, because tasks can obtain more suitable resources more quickly, their execution efficiency, success rate, and likelihood of meeting Service Level Agreements (SLAs) are significantly improved, thereby comprehensively enhancing the quality and reliability of the computing network's services. Attached Figure Description

[0031] Figure 1 This is an exemplary flowchart of a computing power network resource reconstruction method based on multi-element fusion, as shown in some embodiments of this specification.

[0032] Figure 2 This is an exemplary flowchart of a computing power network resource reconstruction method based on multi-element fusion in one embodiment; Detailed Implementation

[0033] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0034] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, an indirect connection through an intermediate medium, or the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0035] Please see Figure 1 In a first aspect, the present invention provides a method for reconstructing computing network resources based on multi-element fusion, comprising:

[0036] Based on the idle resources in the computing power network, a resource list containing four categories of elements is generated;

[0037] Analyze the coupling relationships between the elements in the resource list and generate a resource coupling relationship map;

[0038] Get the task prediction list;

[0039] Based on the resource coupling relationship graph and the task prediction list, a resource reconfiguration strategy is generated;

[0040] In response to state changes in the computing network, resource reconfiguration operations are triggered and executed.

[0041] This method integrates the entire process of resource inventory generation, coupling relationship analysis, task prediction, strategy generation, and dynamic triggering execution, thereby achieving unified perception, deep understanding, and forward-looking optimization of multi-dimensional resources of computing power networks.

[0042] Its core benefits lie in: firstly, the systematic integration and analysis of four key elements—computing power, algorithms, data, and models—and resource modeling based on resource coupling relationship graphs, enabling the quantification and utilization of the inherent synergistic effects between resources; secondly, the introduction of a task prediction mechanism based on historical data, transforming resource reconfiguration from a passive response to an active adaptation, significantly improving preparedness for future task demands; and thirdly, by real-time monitoring of resource and task status and setting trigger conditions, ensuring that reconfiguration strategies can be dynamically updated and executed according to the network environment, enhancing the system's real-time performance and adaptability. This method fundamentally improves the intelligence level and overall efficiency of computing power network resource management.

[0043] One way to generate a resource list is to periodically scan the resource registration information of each node in the computing power network, extract the resource type (such as CPU / GPU computing power, specific algorithm library, dataset, model file) and its attribute parameters (such as computing power floating-point performance, algorithm complexity, data format, model parameter quantity) from the resource description file, and store them in a structured manner according to the preset four-element classification template to form a resource list.

[0044] Another way to generate a resource list is to receive resource metadata in real time through a standardized resource description interface when a resource provider registers new resources with the computing power network, and update the list to the central resource list database in an instant to ensure the real-time nature of the list.

[0045] Those skilled in the art will understand that the resource list can also be generated through other means such as reporting by proxy nodes and aggregation of resource catalog services. As long as the four types of resource information can be collected and classified comprehensively and accurately, they should all fall within the protection scope of this invention.

[0046] In some embodiments, according to the method of the first aspect, generating the resource list includes:

[0047] Identify and record the types and performance indicators of computing resources;

[0048] Identify and record the types and execution characteristics of algorithm resources;

[0049] Identify and record the format and distribution characteristics of data resources;

[0050] Identify and record the size and structural parameters of the model resources.

[0051] By refining and standardizing the recording of four types of resource elements, a precise data foundation is laid for subsequent coupled analysis. Recording the characteristics of each type of resource (such as computing power type, algorithm complexity, data distribution, and model size) makes the resource description more comprehensive and structured, facilitating accurate system assessment of resource capabilities and application scenarios.

[0052] For computing resources, performance metrics may include, but are not limited to: peak computing power (FLOPS), memory capacity and bandwidth, storage I / O performance, and network latency and bandwidth. For example, computing performance data can be obtained by running standard benchmark programs (such as Linpack and HPL).

[0053] For algorithm resources, execution characteristics may include, but are not limited to: time complexity, space complexity, average execution latency, parallelization support, and input / output data format requirements. This information can be obtained from the metadata of the algorithm library or through lightweight performance profiling.

[0054] Those skilled in the art will understand that the above-mentioned records can be expanded or adjusted according to the specific computing power network environment, such as adding power consumption indicators, licensing agreement information, etc., all of which are reasonable extensions of the present invention.

[0055] In some embodiments, according to the method of the first aspect, generating the resource coupling relationship graph includes:

[0056] Analyze the compatibility between data and algorithms;

[0057] Analyze the relationship between computing power and algorithm execution efficiency;

[0058] Analyze the synergistic effect between the algorithm and the model;

[0059] Based on the aforementioned relationship, a graph is constructed with resources as nodes and coupling degree as edge weights.

[0060] By quantitatively analyzing the interactions between three core pairs of elements—data-algorithm, computing power-algorithm, and algorithm-model—the synergistic potential between resources can be accurately characterized. Transforming the analysis results into a weighted graph model makes complex multi-element coupling relationships visible and computable, providing an intuitive and efficient decision-making basis for finding the optimal resource combination. This graph-based representation method is particularly suitable for handling complex correlation problems between large-scale, heterogeneous resources.

[0061] Analyze and evaluate the coupling relationship between data resources and algorithm resources, construct an algorithm requirement matrix to analyze whether the data can meet the input requirements for algorithm execution. If so, test the accuracy and efficiency of the algorithm on different datasets to quantify the algorithm's adaptability to the data.

[0062] One approach to analyzing data-algorithm fit is to define the required input data pattern for each algorithm, calculate the matching degree against the metadata description of the data resources, and use the matching degree score as part of the coupling degree. Furthermore, the algorithm can be run on a sample dataset for testing, and the appropriate score can be fine-tuned based on performance metrics such as accuracy and recall.

[0063] The coupling relationship between computing power resources and algorithm resources is analyzed and evaluated. Based on the algorithm's complexity and benchmark tests, the execution time of the algorithm on different computing power is evaluated, and a curve showing the relationship between algorithm execution time and computing power requirements is constructed. For algorithms executed under the same computing power conditions, the shorter the execution time, the higher the degree of coupling.

[0064] One approach to analyzing computing power and algorithm execution efficiency is to deploy benchmark tests of algorithms on heterogeneous computing power nodes, record their execution time, throughput, and other metrics under different computing power configurations, construct a "computing power configuration-performance" mapping table, and normalize the performance into an efficiency score as the coupling degree.

[0065] Analyze and evaluate the coupling relationship between algorithm resources and model resources, and test the convergence speed and accuracy performance of different algorithms on the same model. Construct a mapping relationship between algorithm complexity and model size to quantify the coupling strength; under the same model size, the lower the algorithm complexity, the higher the coupling.

[0066] By integrating the coupling results between various elements, each element resource is regarded as a node. If the resource coupling of each element reaches a threshold, a bidirectional edge is connected between the two resources. The weight of the edge is the degree of coupling between the resources, thereby generating a resource coupling relationship graph, which provides guidance for resource reconstruction.

[0067] Those skilled in the art will understand that the analysis of coupling relationships is not limited to the above three pairs. It is also possible to add the analysis of data-model (such as the impact of data quality on model accuracy), computing power-model (such as the impact of computing power on model inference speed), and other relationships to construct more complex hypergraph models. All of these variations should fall within the protection scope of this invention.

[0068] In some embodiments, according to the method of the first aspect, obtaining the task prediction list includes:

[0069] Collect historical mission information;

[0070] Predicting the demand characteristics of future tasks based on machine learning models;

[0071] Generate a prediction list that includes task type, resource requirements, and constraints.

[0072] By analyzing historical task data and using machine learning for prediction, the system can anticipate future load trends and resource demand patterns. This proactive capability allows the system to pre-optimize or pre-allocate resource structures before tasks actually arrive, thereby reducing task waiting time, smoothing resource load, and significantly improving service quality and resource utilization.

[0073] One way to achieve this prediction is to use time series forecasting models (such as ARIMA and LSTM) to model the arrival time, quantity, and type of historical tasks, and then predict the arrival patterns of tasks in future periods.

[0074] Another prediction approach is to incorporate task context features (such as user type and business scenario labels) in addition to time series data, and use classification or regression models (such as random forests and gradient boosting trees) to predict the specific demand of future tasks for four types of resources (such as how much GPU computing power is needed and what type of pre-trained model is required).

[0075] Those skilled in the art will understand that the selection of prediction models and feature engineering can be adjusted according to specific application scenarios, and the prediction results of multiple models can be integrated to improve the robustness and accuracy of predictions.

[0076] In some embodiments, according to the method of the first aspect, the triggering and execution of the resource refactoring operation includes:

[0077] Monitor changes in resource status, and when new resources reach a threshold, update the resource list and resource coupling relationship graph, and regenerate the strategy;

[0078] The system monitors the arrival of new tasks, assesses the coupling between the new tasks and existing resources, and triggers a refactoring process if the degree of coupling is below a threshold.

[0079] Based on the current resource restructuring strategy, resources are decomposed and reorganized to form a resource combination adapted to the new task.

[0080] This feature establishes a clear and automated refactoring trigger mechanism to ensure the system can respond promptly to changes in the internal and external environment. Through dual-path monitoring (resource changes and task arrival), both the timeliness of the resource model and the adaptability of task execution are guaranteed. Once a refactoring is triggered, the system can automatically execute complex resource disassembly and reorganization processes based on the latest strategies, dynamically orchestrating the static resource pool into functional units that meet specific task requirements, thus achieving "on-demand resource construction."

[0081] One scenario where resource status changes trigger reconstruction is when the number of newly registered GPU servers monitored reaches a preset threshold (e.g., 5 servers). The system automatically adds their information to the resource list, recalculates their coupling relationship with existing algorithms and models, updates the resource coupling relationship graph, and may generate new reconstruction strategies that can utilize the computing power of these newly added GPUs (e.g., building a distributed training cluster).

[0082] One scenario where task arrival triggers reconfiguration is when a newly arrived image recognition task requires a specific convolutional neural network model and a high-performance GPU. The system queries the resource coupling graph to find currently idle GPUs and...

[0083] The required model has a high degree of coupling, but insufficient coupling with the optimal inference algorithm. At this point, the system triggers a refactoring, and according to the strategy, "schedules" or "instantiates" the highly coupled and efficient inference algorithm from another node to that GPU node, combining it with the model to form a complete service unit.

[0084] Those skilled in the art will understand that the threshold can be a fixed value or a dynamically adaptive one (e.g., learned based on historical change frequencies). For example, the resource quantity threshold can be dynamically adjusted according to a certain proportion of the current total resources; the task coupling threshold can be adaptively set based on the success rate of historical task matching. The reconfiguration operation can be a direct reconfiguration of physical resources, or it can be a resource reorganization at the logical layer through virtualization and containerization technologies.

[0085] Secondly, the present invention provides a computing power network resource reconstruction system based on multi-element fusion, comprising:

[0086] The resource inventory management unit is used to generate and maintain a resource inventory that includes four categories of elements: computing power, algorithms, data, and models.

[0087] The coupling analysis unit is used to analyze the coupling relationships between resource elements and generate a resource coupling relationship map;

[0088] The task prediction unit is used to generate a task prediction list.

[0089] The strategy generation unit is used to generate a resource reconstruction strategy based on the resource coupling relationship graph and the prediction list;

[0090] The monitoring and execution unit is used to trigger and control the resource reconfiguration process in response to changes in the state of resources or tasks.

[0091] This system implements all the functions of the above methods using a modular architecture, with clear responsibilities and collaborative work among the various units. The resource inventory management unit ensures a unified view of all network resources; the coupling analysis unit enables in-depth mining of the inherent relationships between resources; the task prediction unit provides the system with foresight; the strategy generation unit is the core of the system's decision-making; and the monitoring and execution unit ensures the real-time implementation of decisions. This architecture gives the system good scalability, maintainability, and robustness, enabling deployment in large-scale computing network environments.

[0092] The various units of the system can be deployed on the control plane of the computing network as a centralized manager; or they can adopt a distributed architecture, with each unit implemented by multiple cooperating microservices to improve reliability and scalability.

[0093] In some embodiments, according to the system of the second aspect, the resource inventory management unit includes:

[0094] The computing power identification module, algorithm identification module, data identification module, and model identification module are used to identify and record the attribute information of the corresponding categories of resources, respectively.

[0095] By establishing dedicated identification modules to handle different types of resources, the accuracy and efficiency of resource information collection have been improved. Each module can optimize its identification protocol and parsing logic for specific resource types. For example, the computing power identification module may call the hardware information interface, while the algorithm identification module may parse the metadata of the code repository.

[0096] In some embodiments, according to the system of the second aspect, the coupling analysis unit includes:

[0097] The data-algorithm coupling analysis module, the computing power-algorithm coupling analysis module, and the algorithm-model coupling analysis module are used to quantify the coupling strength between different elements, respectively.

[0098] The graph construction module is used to integrate analysis results to generate a weighted resource coupling relationship graph.

[0099] By decomposing complex coupled analysis tasks into multiple specialized sub-modules for parallel processing, analysis efficiency is improved. The graph construction module is responsible for normalizing and fusing the quantization results from each sub-module and constructing a unified graph data structure, providing standardized input for upper-level strategy generation.

[0100] In some embodiments, the monitoring and execution unit includes:

[0101] The resource monitoring module is used to monitor resource registration and updates, trigger resource list and graph updates when conditions are met, and regenerate the reconstruction strategy.

[0102] The task monitoring module is used to monitor the arrival of new tasks and assess their compatibility with existing resources; the refactoring triggering and scheduling module is used to invoke the resource refactoring strategy and perform resource reorganization when the compatibility is insufficient.

[0103] This unit enables dynamic, separate monitoring and unified response to the "resource side" and "task side". The resource monitoring module ensures the freshness of the resource model, the task monitoring module ensures the feasibility of task execution, and the reconstruction triggering and scheduling module is the "execution engine" that links the monitoring results with the reconstruction strategy, and is responsible for coordinating the specific operations of resource reorganization (such as starting containers, configuring networks, mounting storage, etc.).

[0104] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0105] The method of this invention is solidified into a widely distributed and deployable software product via a storage medium. Users or system integrators can load and run this program to equip their computing devices with the aforementioned advanced computing network resource reconfiguration capabilities, thereby facilitating the upgrade or construction of an intelligent computing network management platform.

[0106] See Figure 2 Taking a large-scale AI model training task as an example, the implementation process of the method of the present invention will be explained in detail.

[0107] Resource Inventory Generation: The system scans the computing power network and discovers a large number of scattered small servers with idle GPUs (intelligent computing power) that cannot independently train large-scale AI models; as well as training datasets and several distributed training algorithms stored on multiple nodes. The resource inventory management unit categorizes and records this information to form an initial inventory. The recorded content includes: specific information such as the computing power, data transmission rate, and memory utilization of each computing resource; the complexity and latency of the training algorithms; the distribution characteristics and storage format of the data resources; and the scale and parameter requirements of the analyzed models.

[0108] Resource coupling relationship graph construction: The coupling analysis unit starts working, analyzing and evaluating the coupling relationships between data resources and algorithm resources, computing power resources and algorithm resources, and algorithm resources and model resources in the resource list. It integrates the coupling results between various elements, treats each element resource as a node, and if the resource coupling of each element reaches a threshold, a bidirectional edge is connected between the two resources. The weight of the edge is the degree of coupling between the resources, thereby generating a resource coupling relationship graph to provide guidance for resource reconstruction.

[0109] The data-algorithm analysis module tests the data loading efficiency and preprocessing compatibility of different distributed training algorithms on existing datasets. The computing power-algorithm analysis module tests the single-machine performance and scalability of these algorithms on small GPU servers. The algorithm-model analysis module evaluates the convergence speed of algorithms training specific large models. Finally, the graph construction module generates a resource coupling graph, showing the high coupling weights between certain distributed training algorithms and the current data format, small GPU computing power combinations, and large model structures.

[0110] Task Prediction: Based on historical task requests, the task prediction list generates a list of training tasks that may be submitted in the near future, including their size, resource requirements, and data distribution characteristics. This includes large-scale AI model training tasks.

[0111] For example, the task prediction unit analyzes historical logs and discovers that large-scale model training tasks are frequently submitted at the beginning of each week. Based on an LSTM model, it predicts that new training tasks are likely to arrive within the next 24 hours, with the following characteristics: large-scale parallel training required, specific types of datasets, and a huge number of parameters in the target model. Based on this, a task prediction list is generated.

[0112] Strategy Generation: The strategy generation unit analyzes the resource coupling graph and the prediction list. It discovers that the predicted task cannot be satisfied by any single server, but the resource coupling graph shows that multiple small GPU servers and scattered datasets can be combined to work collaboratively through an efficient distributed training algorithm. Specifically, the strategy generation unit queries the resource coupling graph and, based on the high-weight edges shown, identifies a series of small GPU computing power nodes that have historically had high coupling with the distributed training algorithm and low network latency among them, as a candidate reorganization set. Therefore, it generates a reorganization strategy: logically aggregating several small GPU servers into a virtual training cluster and assigning it the distributed training algorithm and data scheduling scheme with the highest coupling.

[0113] Monitoring and triggering: Real-time monitoring and updating of resources. When a certain number of new resources are registered, the resource list and coupling relationship graph are updated to ensure that the resource reconstruction strategy is based on the latest data.

[0114] For example, the resource monitoring module detects the registration of two new high-performance GPU servers during this period. When the number of new resources reaches a preset threshold (e.g., a cumulative total of three new servers; in this example, two servers did not reach the threshold, but the system may determine the need for an update based on a dynamic threshold), it triggers an update to the resource list and resource coupling graph. The task monitoring module then detects that the predicted large-scale training task has actually arrived. An assessment shows that the coupling between the existing resource structure (independent servers) and the task is below the preset threshold.

[0115] Refactoring Execution: The refactoring triggering and scheduling module is activated, invoking the latest refactoring strategy. The system automatically performs the following operations: selecting a set of optimal small servers and newly registered high-performance servers according to the strategy; deploying the specified distributed training algorithm container on the selected servers; configuring a distributed file system or data caching service for efficient access to the distributed dataset; initializing the parameters of the large model to be trained and distributing them to each computing node. Finally, a temporarily constructed, highly adaptable distributed training environment is ready, and the task begins execution.

[0116] Through this embodiment, the computing network successfully reconstructs originally idle and scattered heterogeneous resources into a whole that can efficiently process complex tasks, fully demonstrating the significant effect of the present invention in improving resource utilization and enhancing system adaptability.

[0117] This embodiment uses real-time face recognition analysis in an urban smart security scenario as an example to illustrate in detail the implementation process of the method of the present invention in an edge computing network. In this scenario, multiple high-definition cameras are deployed at various entrances and exits, and the generated video streams need to be detected and recognized in real time at the edge, and compared with the central database. The task is characterized by low latency and high concurrency.

[0118] S1: Resource Inventory Generation

[0119] The system scans the edge computing network to identify and classify currently available resources:

[0120] Computing resources: Identify edge servers, smart gateways (with built-in NPU), and a small amount of idle cloud GPU resources located in each region, and record their computing power (TOPS, FLOPS), memory size, and current load rate.

[0121] Algorithm Resources: Identify various face detection algorithms (such as MTCNN, YOLO-face) and face recognition algorithms (such as FaceNet, ArcFace) deployed in the algorithm repository, and record their accuracy, latency, model size, and requirements for input image resolution.

[0122] Data resources: Identify real-time video stream data (RTSP / H.264 format) generated by each camera, historical captured image library, and central facial feature database, and record their data format, storage location, access latency, and update frequency.

[0123] Model Resources: Identify various pre-trained face recognition models (such as models optimized for Asian faces and models for recognizing faces while wearing masks), and record their versions, number of parameters, feature dimensions, and applicable scenarios.

[0124] The system integrates the above information to generate a structured resource list.

[0125] S2: Construction of Resource Coupling Relationship Graph

[0126] The coupling analysis unit performs correlation analysis on the resources in the inventory:

[0127] Data-algorithm coupling analysis: This study tested the parsing efficiency and accuracy of different face detection algorithms for real-time video streams (at different resolutions and bitrates); it also tested the recognition accuracy of different face recognition algorithms for face images extracted from the video. Quantitative analysis showed that Algorithm A had the highest adaptability to the video stream from camera 1.

[0128] Computing power-algorithm coupling analysis: The above algorithms were deployed on various edge servers, and their frame rate (FPS) and power consumption were tested. It was found that Algorithm B had the best energy efficiency on Jetson devices, while Algorithm C was the fastest on cloud GPUs but had higher latency.

[0129] Algorithm-model coupling analysis: The combined effects of different recognition algorithms and various face recognition models were tested to evaluate their overall recognition accuracy and processing time. It was found that the combination of model D and algorithm B achieved the lowest latency while maintaining high accuracy.

[0130] Graph Generation: Using the above resources as nodes, if the fit / efficiency score exceeds a threshold (e.g., 80%), weighted edges (weight = score) are established between nodes. The final resource coupling relationship graph clearly displays multiple highly coupled relationships such as: {Camera 1 video stream → Algorithm A}, {Algorithm B → Jetson device}, and {Algorithm B + Model D}.

[0131] S3: Task Prediction List Generation

[0132] The task prediction unit analyzes historical task data (such as the number of facial recognition requests and event types at different times over the past month) and combines it with external data such as weather forecasts and holiday information to use a time series model for prediction. The prediction results show that from 6:00 PM to 8:00 PM on a future Friday, the number of facial recognition requests in the shopping mall area will increase by 300% due to promotional activities, with extremely high real-time requirements (latency <200ms). Based on this, a task prediction list is generated, clearly indicating that "high-concurrency, low-latency facial recognition" task demands will occur in specific areas during this period.

[0133] S4: Resource Reconfiguration Strategy Generation

[0134] The strategy generation unit combines a resource coupling relationship graph with a task prediction list for analysis.

[0135] Prediction tasks require extremely high processing throughput and low latency, and existing statically allocated resources may be insufficient or fail to meet latency requirements.

[0136] The analysis revealed that the video streams from the cameras in the shopping mall area were highly coupled with Algorithm A, Algorithm B was highly coupled with the Jetson device, and the combination of Algorithm B and Model D had the best performance.

[0137] However, these highly coupled resources are currently scattered across different physical nodes, and direct access may introduce delays due to network transmission.

[0138] Therefore, a dynamic resource reconfiguration strategy is generated: before the predicted peak period, the combination of Algorithm A and Algorithm B + Model D is proactively pre-deployed to the edge Jetson devices in the shopping mall area to form a localized processing pipeline of "video stream - Algorithm A - Algorithm B - Model D", and additional computing buffers are pre-allocated.

[0139] S5: Real-time monitoring and resource updates

[0140] The resource monitoring module continued to run. On Friday afternoon, two new Jetson Nano devices were detected registering to the edge network in the shopping mall area. Since the new resources reached a preset threshold (e.g., "≥2 new available high-efficiency edge computing units"), the system automatically triggered a resource list update, adding the new devices to the list and recalculating their coupling relationships with existing algorithms and models, updating the resource coupling relationship graph. Based on the new graph, the policy generation unit fine-tuned the policies in S4, also including the new devices in the pre-deployment plan to further enhance processing capabilities.

[0141] S6: Real-time monitoring and task triggering

[0142] The task monitoring module detected a sharp increase in the request rate of the facial recognition API in the shopping mall area at 18:00 on Friday evening, consistent with predictions. The system immediately assessed the coupling between the currently arriving real-time tasks and the existing resource configuration:

[0143] Currently, most requests still require remote calls to the algorithms and models in the central cloud, and the coupling degree is assessed as "low" (high latency, high bandwidth consumption).

[0144] The coupling degree is lower than the system's set performance threshold (this coupling degree reflects performance metrics including end-to-end latency; for example, the current latency is >500ms, which does not meet the <200ms requirement). Therefore, the system determines that resource reconfiguration is needed to adapt to the burst of task demands and immediately triggers the reconfiguration process.

[0145] S7: Resource Restructuring Execution

[0146] The refactoring triggering and scheduling module is activated, invoking the currently updated resource refactoring strategy:

[0147] Decomposition: Decompose the "localized processing pipeline" specified in the strategy into specific operation instructions: pull and instantiate the container images of algorithm A and algorithm B on the specified Jetson devices (including the two newly added ones); load the parameter file of model D; configure the network channel for direct input of camera video streams to these containers.

[0148] Recombination: The computing power (Jetson device), algorithms (A and B), data (camera video stream), and models (D) are rapidly combined at the edge into a distributed, high-performance processing unit.

[0149] Routing and scheduling: Dynamically schedule newly arriving face recognition requests to newly built edge processing units, rather than remote cloud.

[0150] Within minutes, the system completed dynamic resource reorganization. After reorganization, the face recognition task was completed directly at the edge, reducing end-to-end latency to below 100ms, while alleviating the pressure on core network bandwidth and cloud load, and efficiently handling peak task demands.

[0151] This embodiment demonstrates how, in dynamically changing edge scenarios, the present invention achieves proactive and dynamic resource optimization through a closed loop of "prediction-monitoring-triggering-reconfiguration." The system no longer statically allocates or passively responds, but rather, based on a deep understanding of the coupling relationships between multiple elements and accurate prediction of future needs, it proactively plans and adjusts resource configurations in real time. This prepares the system before peak task periods and rapidly executes the optimal configuration when tasks arrive, significantly improving the service quality, resource utilization, and overall responsiveness of edge computing networks for intelligent applications with extremely high real-time requirements.

Claims

1. A method for reconstructing computing network resources based on multi-element fusion, characterized in that, include: Based on the idle resources in the computing power network, a resource list containing four categories of elements is generated; Analyze the coupling relationships between the elements in the resource list and generate a resource coupling relationship map; Get the task prediction list; Based on the resource coupling relationship graph and the task prediction list, a resource reconfiguration strategy is generated; In response to state changes in the computing network, resource reconfiguration operations are triggered and executed.

2. The method according to claim 1, characterized in that, The generated resource list includes: Identify and record the types and performance indicators of computing resources; Identify and record the types and execution characteristics of algorithm resources; Identify and record the format and distribution characteristics of data resources; Identify and record the size and structural parameters of the model resources.

3. The method according to claim 1, characterized in that, The generated resource coupling relationship graph includes: Analyze the compatibility between data and algorithms; Analyze the relationship between computing power and algorithm execution efficiency; Analyze the synergistic effect between the algorithm and the model; Based on the aforementioned relationship, a graph is constructed with resources as nodes and coupling degree as edge weights.

4. The method according to claim 1, characterized in that, The task prediction list includes: Collect historical mission information; Predicting the demand characteristics of future tasks based on machine learning models; Generate a prediction list that includes task type, resource requirements, and constraints.

5. The method according to claim 1, characterized in that, The triggering and execution of the resource reconfiguration operation includes: Monitor changes in resource status, and when new resources reach a threshold, update the resource list and map, and regenerate the strategy; The system monitors the arrival of new tasks, assesses the coupling between the new tasks and existing resources, and triggers a refactoring process if the degree of coupling is below a threshold. Based on the current resource restructuring strategy, resources are decomposed and reorganized to form a resource combination adapted to the new task.

6. A computing power network resource reconfiguration system based on multi-element fusion, characterized in that, include: The resource inventory management unit is used to generate and maintain a resource inventory that includes four categories of elements: computing power, algorithms, data, and models. The coupling analysis unit is used to analyze the coupling relationships between resource elements and generate a resource coupling relationship map; The task prediction unit is used to generate a task prediction list. The strategy generation unit is used to generate a resource reconstruction strategy based on the graph and the prediction list; The monitoring and execution unit is used to trigger and control the resource reconfiguration process in response to changes in the state of resources or tasks.

7. The system according to claim 6, characterized in that, The resource list management unit includes a computing power identification module, an algorithm identification module, a data identification module, and a model identification module, which are used to identify and record the attribute information of corresponding categories of resources.

8. The system according to claim 6, characterized in that, The coupling analysis unit includes: a data-algorithm coupling analysis module, a computing power-algorithm coupling analysis module, and an algorithm-model coupling analysis module, which are used to quantify the coupling strength between different elements, respectively. The graph construction module is used to integrate analysis results to generate a weighted resource coupling relationship graph.

9. The system according to claim 6, characterized in that, The monitoring and execution unit includes: The resource monitoring module is used to monitor resource registration and updates, trigger resource list and graph updates when conditions are met, and regenerate the reconstruction strategy. The task monitoring module is used to monitor the arrival of new tasks and assess their compatibility with existing resources. The refactoring triggering and scheduling module is used to invoke the resource refactoring strategy and perform resource reorganization when the adaptability is insufficient.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 5.

Images