A cloud-edge-end collaborative scheduling method and system for a computing power network

Abstract

This invention discloses a cloud-edge-device collaborative scheduling method and system for computing power networks, relating to the field of cloud-edge collaborative computing. The method includes: constructing a heterogeneous computing resource characteristic map based on the topological relationship of the target computing power network and the initialization of a graph neural network; embedding and updating the map by combining node state parameters obtained from the cloud-edge-device generalized perception layer; standardizing and decomposing real-time business requirements based on a set of business requirements to obtain real-time business components; obtaining the multi-dimensional scheduling strategy weight coefficients for the target scenario; solving for the optimal route based on multi-objective optimization by combining the real-time business components and the heterogeneous computing resource characteristic map; constructing a multi-dimensional computing power spectrum for the subnet; allocating edge-device tasks based on the real-time business components; and fusing the optimal computing route with the edge-device task allocation results to perform collaborative scheduling of real-time business requirements. This method solves the problems of poor real-time performance of node states, mismatch between resource requirements and node capabilities, and poor optimization results.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaborative computing, specifically to a cloud-edge-device collaborative scheduling method and system for computing power networks. Background Technology

[0002] The computing network connects heterogeneous computing resources in cloud data centers, edge nodes, and terminal devices through the network, forming a three-level collaborative computing service system of cloud, edge, and terminal.

[0003] However, existing computing power network scheduling methods lack a unified performance representation and comparison benchmark among computing power nodes, and the node status changes rapidly, resulting in poor real-time performance of node status; business requirements are usually described in an unstructured form, lacking a standardized quantitative decomposition mechanism, resulting in a mismatch between business resource requirements and node capabilities; network routing and computing task allocation exist independently, resulting in poor optimization effects; and the heterogeneous computing power of edge devices is not fully utilized and lacks synergy. Summary of the Invention

[0004] This application provides a cloud-edge-device collaborative scheduling method and system for computing power networks, which addresses the problems of poor real-time node status, mismatch between business resource requirements and node capabilities, and poor optimization effect of cloud-edge-device task scheduling in existing technologies.

[0005] In view of the above problems, this application provides a cloud-edge-device collaborative scheduling method and system for computing power networks.

[0006] Firstly, this application provides a cloud-edge-device collaborative scheduling method for computing power networks, the method comprising: Based on the topological relationship of the target computing power network, an initial heterogeneous computing resource characteristic map based on graph neural network is established, and the node state parameters obtained by the cloud-edge-device general perception layer are embedded and updated. Obtain real-time business requirements and, in conjunction with a pre-built set of standard business requirements, perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components. Obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, and combine the real-time service components with the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization and obtain the optimal computing route; Based on the optimal computing route and the heterogeneous computing resource characteristic map, a multi-dimensional computing power spectrum of the subnet is constructed, and edge-end task allocation is performed in conjunction with the real-time service components. By integrating the optimal computational route with the edge-to-end task allocation results, collaborative scheduling of real-time business requirements is performed.

[0007] Secondly, the present invention provides a cloud-edge-device collaborative scheduling system for computing power networks, the system comprising: The embedding update module is used to initialize and establish a heterogeneous computing resource characteristic map based on the topological relationship of the target computing power network, and to perform embedding update in combination with the node state parameters obtained by the cloud-edge-device general perception layer. The business requirement decomposition module is used to obtain real-time business requirements and, in conjunction with a pre-built set of standard business requirements, to perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components. The optimal route solving module is used to obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, and combine the real-time service components with the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization to obtain the optimal computing route. The task allocation module is used to construct a multi-dimensional computing power spectrum of the subnet based on the optimal computing route and the heterogeneous computing resource characteristic map, and to perform edge-end task allocation in combination with the real-time service components. The business demand scheduling module is used to integrate the optimal computing route with the edge-end task allocation results to perform real-time collaborative scheduling of business demands.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first establishes a heterogeneous computing resource characteristic map based on topological relationships, and then embeds and updates it by combining node state parameters. This merges the static heterogeneous computing capabilities with the dynamic node states into a structured feature embedding vector, providing a resource foundation for subsequent scheduling decisions. Second, it acquires real-time business requirements and performs standardized requirement decomposition using a standard business requirement set. This transforms business requests into multi-dimensional computing power requirement vectors and business constraint vectors, achieving a mapping between business language and resource language, providing standardized input for subsequent processing. Third, it acquires multi-dimensional scheduling strategy weight coefficients, solves for optimal routing based on multi-objective optimization to obtain the optimal computing route, and adaptively adjusts the weights according to scenario preferences to find the optimal cloud-edge computing path globally. Furthermore, it constructs a subnet multi-dimensional computing power spectrum based on the optimal computing route and the heterogeneous computing resource characteristic map, and performs edge computing... End-to-end task allocation focuses global routing within the edge subnet, enabling node matching on the edge and resource allocation from macroscopic paths to microscopic nodes. Ultimately, it integrates optimal computational routing with edge... The task allocation results are used for collaborative scheduling to achieve efficient, reliable, and adaptive collaborative scheduling for real-time business needs. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a cloud-edge-device collaborative scheduling method for computing power networks according to this application; Figure 2This is a schematic diagram of the structure of a cloud-edge-device collaborative scheduling system for computing power networks according to this application.

[0010] In the attached diagram, the components represented by each number are as follows: Embedded update module 11, business requirement decomposition module 12, optimal route solving module 13, task allocation module 14, business requirement scheduling module 15. Detailed Implementation

[0011] This application provides a cloud-edge-device collaborative scheduling method for computing power networks, which specifically addresses the problems of poor real-time node status, mismatch between business resource requirements and node capabilities, and poor optimization effect of cloud-edge-device task scheduling in existing technologies.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides a cloud-edge-device collaborative scheduling method for computing power networks, the method comprising: S10: Based on the topological relationship of the target computing power network, initialize and establish a heterogeneous computing resource characteristic map based on graph neural network, and embed and update it by combining the node state parameters obtained by the cloud-edge-device general perception layer. In this embodiment, the target computing power network is a specific computing infrastructure environment for resource scheduling, which is typically formed by interconnecting cloud center nodes, edge nodes, and terminal device nodes through a communication network; the topology relationship is the physical or logical connection structure and connection distance between various computing power nodes in the target computing power network, which is the basic data used to describe network connectivity; the graph neural network is a deep learning model specifically for processing graph structure data, which updates the feature representation of the current node by iteratively aggregating the feature information of neighboring nodes, and can effectively capture the topological dependencies in graph data.

[0014] The heterogeneous computing resource characteristic map is a dynamic knowledge base built on a graph structure; the cloud-edge-device general perception layer is a distributed data acquisition and monitoring system that logically spans the cloud, edge, and device levels. Through agent programs deployed on nodes at each level, it is responsible for acquiring the running status of nodes in real time or near real time; node status parameters include, but are not limited to, dynamic change indicators such as the node's central processing unit utilization, memory usage, network input / output throughput, task queue length, energy consumption level, and current available computing power capacity; embedding update is, in the context of graph neural networks, the process of mapping high-dimensional, sparse original node attribute information to a low-dimensional, dense vector space, and iteratively optimizing the embedding representation of the current node.

[0015] Specifically, an initial graph structure is constructed based on the physical or logical topology of the target computing network. Subsequently, a graph neural network is used for initialization to establish a heterogeneous computing resource characteristic graph. In this graph, each node's initial feature vector contains its inherent performance data for different standard computing tasks. Through forward propagation computation, the embedding vectors of all nodes in the graph are iteratively updated in one or more rounds. The updated node embedding vectors retain their static computing power characteristics and also incorporate dynamic state information such as current load and availability, as well as network state information propagated from neighboring nodes.

[0016] Step S10 in the method provided in this application embodiment includes: The cloud-edge-device generalized perception layer is used to perform hybrid perception and acquisition on the target computing network to obtain node status parameters. Based on the aforementioned topological relationship, a graph structure is constructed using the relative performance values ​​of computing power nodes and computing power nodes facing multiple preset standard computing tasks as graph nodes, and the network distance between computing power nodes as relation edges. Based on the node state parameters, the graph nodes and the relation edges are embedded and updated in the graph structure to generate feature embedding vectors of multiple graph nodes, thereby obtaining the heterogeneous computing resource characteristic map. The feature embedding vector is defined as the embedding vector of itself updated by a nonlinear transformation function after weighting the neighbor information of neighboring nodes with edge weights.

[0017] In this embodiment, the cloud-edge-device pervasive perception layer is a three-level distributed data acquisition architecture that logically spans cloud data centers, edge computing nodes, and terminal devices; hybrid perception acquisition is a data acquisition strategy that combines multiple acquisition triggering mechanisms; node status parameters are a set of quantitative indicators that describe the current operating status of computing nodes and are the key basis for judging whether a node is suitable for undertaking computing tasks.

[0018] First, the perception agents pre-deployed in various cloud centers, edge servers, and terminal devices within the target computing power network are activated to form a cloud-edge-device pervasive perception layer. Then, data collection operations are performed on the target computing power network according to a pre-defined hybrid perception and acquisition strategy.

[0019] Specifically, differentiated acquisition methods are adopted for state parameters of different node levels or with different characteristics: for parameters with gradual changes, a longer adaptive period is used for acquisition; for parameters with drastic changes, a shorter period or event-driven method is used for acquisition. Through hybrid sensing acquisition, the latest state parameters of each computing node are continuously obtained, such as the current CPU utilization rate of an edge node being 67%, memory usage rate being 82%, and task queue depth being 5. The original state parameters are then temporarily stored in the local cache of the sensing layer, awaiting further processing.

[0020] Secondly, based on topological relationships, a graph structure is constructed using computing power nodes and their relative performance values ​​for multiple preset standard computing tasks as graph nodes, and network distances between computing power nodes as relational edges. A computing power node is an entity in the computing power network that can provide computing resources, including cloud servers, edge gateways, embedded devices, smartphones, and IoT sensors. Each computing power node corresponds to a vertex in the graph. The multiple standard computing tasks are pre-defined, representative benchmark computing loads. The relative performance value is a performance measure of a computing power node relative to a benchmark node when executing a standard computing task under the same standardized testing environment. Network distance is a comprehensive cost measure at the communication layer between two computing power nodes.

[0021] Specifically, the first step is to obtain the full topology of the target computing power network, including the identifiers, location information, and link information between all computing power nodes. Then, each computing power node is mapped to a vertex in a graph structure, thus being considered a graph node. For each vertex, the initial feature vector is a vector composed of the node's relative performance values ​​for several preset standard computing tasks. For example, if five standard computing tasks are preset, the initial feature vector for each node is a 5-dimensional vector, with each dimension representing the node's normalized performance score relative to a global benchmark.

[0022] The network distance between any two computing nodes is calculated and used as the weight of an undirected or directed edge connecting these two nodes. Finally, after processing all nodes and edges, the results are integrated to output a complete graph structure. In the graph structure, each node carries an initial feature vector, and each edge carries a weight.

[0023] For example, the computing network includes a cloud center node E, and the preset standard computing tasks are: T1: integer operation, T2: floating-point operation. Through offline testing, the normalized relative performance value is obtained, and the relative performance value of node E is (0.6, 0.4). In the constructed graph structure, the feature of node E is [0.6, 0.4].

[0024] Finally, based on node state parameters, the graph nodes and relation edges are embedded and updated within the graph structure, generating feature embedding vectors for multiple graph nodes. This yields a heterogeneous computing resource characteristic map. The feature embedding vector is created by weighting the neighbor information of neighboring nodes with edge weights and then introducing nonlinear capabilities through nonlinear transformation functions such as ReLU to update its own embedding vector. After embedding and updating, each graph node's final vector representation is a feature embedding vector, with dimensions such as 64 or 128. Its value encodes the node's comprehensive state in the global resource map, including its inherent computing power, current load, available resources, and topological relationships with surrounding nodes.

[0025] Specifically, the initial graph structure is taken as input, and real-time node state parameters are used as dynamic features, concatenated into the initial feature vector of each node. Then, a graph neural network model, such as a graph convolutional network (GCN), a graph attention network (GAT), or a graph isomorphic network (GIN), is invoked to perform multiple rounds of message passing and aggregation operations according to the edge connections in the graph structure. In each round, each node collects the current feature embedding vectors of all its neighbors, weights the neighbor information using edge weights, and then updates its own embedding vector through a non-linear transformation function. After several iterations, each node's new embedding vector contains its own static performance and dynamic state, while aggregating information from multi-hop neighbors, ultimately outputting the final feature embedding vectors of all graph nodes, forming a heterogeneous computing resource characteristic map.

[0026] For example, the graph neural network model invoked can be constructed based on the graph convolutional network (GCN), and the specific steps are as follows: The GCN model consists of an input layer, hidden layers, and an output layer stacked together. Each layer takes a node feature matrix and an adjacency matrix as input and outputs an updated node feature matrix. The input layer receives the original graph structure data. Assuming there are N nodes in the computing network, each node's initial feature vector includes: normalized performance representation values ​​for multiple preset standard computing tasks and real-time node state parameters obtained through the cloud-edge-device perceptual layer. This is used to receive the original node features and topology, providing input for subsequent graph convolution. The hidden layer is a graph convolutional layer with an input dimension F and an output dimension of 64, using the ReLU activation function to map the original features to a high-dimensional space and aggregate neighbor information. The output layer is a linear mapping layer with an input dimension of 32 and an output dimension of 16, generating the final feature embedding vector Z = N × 16 for each node.

[0027] An unsupervised training approach is employed, with the loss function composed of a graph reconstruction loss and a state prediction loss. The graph reconstruction loss is based on the cross-entropy of the reconstructed adjacency matrix using the embedding inner product, while the state prediction loss is derived from the mean squared error of the node state parameters through two additional MLP layers. The two losses are then weighted and summed. The Adam optimizer is used during training with an initial learning rate of 0.001 and a weight decay of 5e. -4 Weights are uniformly initialized, with a maximum of 500 iterations and a batch size equal to the total number of nodes in the graph, N. Forward propagation calculates the loss layer by layer. Backpropagation calculates the gradient of each layer's weights using automatic differentiation and updates parameters using Adam. After training, a validation set not used in training is used to calculate the link prediction accuracy and the mean squared error of state prediction. The validation loss is considered complete when it fails to decrease for 30 consecutive iterations or its absolute value is less than 1e^(-1 / 2). -5 Stop training and save the best model.

[0028] For example, after 120 training rounds, the validation loss decreased to 0.35, the link prediction accuracy reached 0.92, the mean square error of state prediction was 0.028, and the embedding stability cosine similarity was 0.94, indicating that the model has passed validation. The generated 16-dimensional embedding vector constitutes a heterogeneous computing resource characteristic map.

[0029] In step S10 of the method provided in this application embodiment, the target computing power network is subjected to hybrid sensing acquisition through the cloud-edge-device generalized sensing layer to obtain node state parameters, including: The cloud-edge-device pervasive perception layer includes a cloud center perception agent, an edge-side perception agent, and a device-side perception agent. The hybrid perception acquisition includes event-driven acquisition and adaptive periodic acquisition. The cloud center perception agent and the edge side perception agent perform adaptive periodic data collection, wherein the collection period of the adaptive periodic data collection is negatively correlated with the historical state change rate under the edge side computing power node. The edge-side sensing agent and the endpoint-side sensing agent synchronously perform the adaptive periodic acquisition and the event-driven acquisition.

[0030] In this embodiment, the cloud-edge-device general perception layer firstly includes a cloud center perception agent, an edge-side perception agent, and a device-side perception agent, and hybrid perception collection, including event-driven collection and adaptive periodic collection. The cloud center perception agent is a software component or service process deployed inside the cloud data center, used to periodically collect status parameters from cloud computing power nodes such as cloud servers, virtual machines, and containers, and has high computing and storage resources.

[0031] Edge-side perception agents are software agents deployed on edge computing nodes. They are responsible for collecting the edge node's own state parameters, communicating with its subordinate end-side perception agents, and aggregating data. The resource capabilities of edge-side perception agents fall between those of the cloud center and the end-side. End-side perception agents are extremely lightweight software modules deployed on terminal devices. Because end-side devices are typically resource-constrained, end-side perception agents are designed to be minimally resource-intensive and feature-simplified components, primarily responsible for collecting local state parameters.

[0032] Event-driven acquisition is a data acquisition mechanism triggered by specific events occurring on the monitored object. Events include, but are not limited to: sudden changes in node status parameters exceeding preset thresholds, task completion, node failure or recovery, and changes in network connection status. Adaptive periodic acquisition is a data acquisition mechanism that dynamically adjusts the acquisition time interval. It can automatically shorten or extend the acquisition period based on the historical rate of change or the current dynamic level of the monitored node status.

[0033] Specifically, a cloud-center perception agent is deployed within the cloud data center of the target computing power network, an edge-side perception agent is deployed on each edge node, and an end-side perception agent is deployed on each terminal device. The cloud-center perception agent performs complex data preprocessing and aggregation analysis; the edge-side perception agents collect the state parameters of the edge nodes themselves and communicate and aggregate data with their subordinate end-side perception agents; and the end-side perception agents collect local state parameters. These collaborative processes form a distributed perception system covering the entire network. Furthermore, the hybrid perception acquisition consists of event-driven acquisition and adaptive periodic acquisition. Adaptive periodic acquisition is used to smoothly track general trends in node state changes, maintaining basic data timeliness with controllable overhead; event-driven acquisition is used to capture sudden state changes, ensuring the system can immediately detect anomalies or critical changes.

[0034] Secondly, adaptive periodic data collection is performed between the cloud center perception agent and the edge perception agent. The collection period of the adaptive periodic data collection is negatively correlated with the historical state change rate under the jurisdiction of the edge computing power node. The historical state change rate is a statistical measure of the degree of change of the state parameters of all edge nodes under the jurisdiction of the edge node and the edge node itself over a historical period of time, reflecting the overall dynamic level of the edge subnet.

[0035] Specifically, each edge computing node continuously records its own historical state parameters and those of all its subordinate edge nodes, and calculates the historical state change rate. For example, it calculates the average of the absolute values ​​of the differences between a certain utilization rate and the next utilization rate in the CPU utilization sequence of all its subordinate nodes, as the historical state change rate.

[0036] The data collection cycle is negatively correlated with the historical state change rate: a higher historical state change rate indicates a more unstable state within the edge subnet. Based on this negative correlation principle, the data collection cycle between the cloud and the edge will be automatically shortened, allowing the cloud center to obtain the latest state updates of the edge subnet more frequently. Conversely, if the historical state change rate is low, indicating a stable state within the subnet, the data collection cycle will be automatically extended to save bandwidth resources between the cloud and the edge and processing overhead for the cloud center.

[0037] For example, the CPU utilization sequence of all subordinate nodes is: [0.8, 0.75, 0.9, 0.75, 0.7]. The historical state change rate is calculated as (│0.8-0.75│+│0.75-0.9│+│0.9-0.75│+│0.75-0.7│) / 5=0.08. The reporting cycle from edge node E to the cloud center is automatically shortened to 5 seconds. When the night garden is shut down, the state change rate drops to 0.01, and the reporting cycle is automatically extended to 300 seconds.

[0038] Furthermore, adaptive periodic data collection and event-driven data collection are performed synchronously between the edge-side perception agent and the end-side perception agent.

[0039] Specifically, each edge node, based on its own adaptively adjusted period according to its own state change rate, periodically reports its routine state parameters such as CPU utilization and memory usage to the edge agent. Simultaneously, the edge sensing agent continuously monitors various preset events locally, and immediately triggers additional state reporting upon detecting these events. Periodic data collection ensures continuous updates to baseline data, while event-driven data collection ensures zero-latency detection of sudden anomalies or critical state changes. Upon receiving event-driven reports, the edge agent can immediately update the multi-dimensional computing power spectrum of the subnet it maintains, triggering a real-time re-matching process in the upper-layer scheduling system.

[0040] In this embodiment, the target computing network is subjected to hybrid sensing and acquisition through a cloud-edge-device generalized sensing layer, and a graph structure is constructed based on the topological relationship. Then, the graph neural network is embedded and updated in combination with the node state parameters. This achieves a negative correlation between the acquisition cycle and the rate of change of the historical state under the edge side, providing a unified and computable feature input basis for subsequent multi-objective optimization routing and task allocation.

[0041] S20: Obtain real-time business requirements, and combine them with a pre-built set of standard business requirements to perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components; In this embodiment, real-time service requirements are specific requests initiated by users, applications, or upper-layer orchestrators at runtime that require computing services from the computing power network; the standard service requirement set is the basic computing tasks obtained by decomposing complex service requirements, as well as the typical composition ratio of each service type to the basic tasks; standardized requirement decomposition is the process of mapping the original unstructured real-time service requirements to the normative framework defined in the standard service requirement set; real-time service components are the results obtained after standardized requirement decomposition.

[0042] Upon receiving real-time business requirements, these requirements need to be parsed to match the heterogeneous computing resource characteristic map in subsequent steps. Specifically, a pre-built set of standard business requirements is invoked to identify the current real-time business type. Then, based on the standards for business types in the set of standard business requirements, the original requirements are decomposed into combinations of standard computing tasks, and the specific computing power required for each standard task is determined, resulting in a multi-dimensional computing power requirement. Simultaneously, hard constraints are extracted from the original requirements to form business constraint information. Finally, the multi-dimensional computing power requirement and business constraint information are vectorized and concatenated to generate the real-time business components.

[0043] Step S20 in the method provided in this application embodiment includes: Identify the business type of the real-time business requirement, and invoke business process information based on the business type; By combining the business process information with the standard business requirement set, the real-time business requirements are decomposed into standardized requirements to obtain a multi-dimensional standard task coefficient set. Based on the multidimensional standard task coefficient set and the standard business requirement set, the multidimensional computing power requirement of the real-time business requirement is determined; Based on the business process information and the business type, obtain the corresponding business constraint information, wherein the business constraint information includes at least the upper bound of delay, the lower bound of reliability, and the security level; The multidimensional computing power demand and the business constraint information are vectorized and concatenated to form the real-time business components.

[0044] In this embodiment of the application, the business type of real-time business needs is first identified, and business process information is called based on the business type. The business process information is structured data that describes the standard calculation steps, data flow dependencies, and temporal relationships between steps of a specific business type from input to output.

[0045] Specifically, upon receiving a real-time business request, the system first parses it to identify its corresponding business type. Identification methods can include rule-based keyword matching or semantic understanding based on machine learning classifiers.

[0046] Rule-based keyword matching first pre-constructs a keyword dictionary for each business type. Then, when parsing real-time business requirements, the requirement text is segmented and preprocessed. Next, it matches each business type's keyword dictionary one by one, counting the number of hits, and selecting the business type with the highest hit rate or the first matched type as the recognition result. Semantic understanding based on machine learning classifiers, on the other hand, first constructs an offline labeled dataset, collecting a large amount of historical business requirement text and manually labeling its business type. Each requirement text is converted into a fixed-dimensional semantic vector through word embedding. Then, a classifier is trained, such as a support vector machine, multilayer perceptron, or a fine-tuned pre-trained language model like BERT. The input is the semantic vector, and the output is the probability distribution of the business type. During online recognition, the real-time business requirement text is input into the trained classifier model. The model calculates the confidence score for each type through forward propagation and selects the business type with the highest confidence score as the output to determine the business type.

[0047] After identifying the business type, the identified business type is used as an index to search within a pre-built business process library, retrieving the corresponding business process information. For example, if the identified business type is industrial visual quality inspection, the system's business process information might include the following sequence of steps: image acquisition → image preprocessing → defect detection → classification and counting → result reporting. This business process information provides a structured framework for subsequent resource requirement decomposition.

[0048] Secondly, by combining business process information with a standard business requirement set, real-time business requirements are decomposed into standardized requirements to obtain a multi-dimensional standard task coefficient set. The standard business requirement set is a pre-built knowledge base, in which the proportion coefficient of each standard computing task in the typical execution process is defined for each business type. Standardized requirement decomposition is the process of mapping business flows described in general terms to standard computing tasks with quantifiable performance indicators in the standard business requirement set. The multi-dimensional standard task coefficient set is a set of coefficients for standard tasks in multiple dimensions, where the dimension is equal to the total number K of predefined standard computing tasks in the standard business requirement set.

[0049] Specifically, the system receives business process information and simultaneously reads a pre-built set of standard business requirements. This set records typical task coefficient templates for each standard business type. The current business process information is compared with the standard templates for that business type. Since actual business operations may have parameter differences, such as resolution differences, the coefficients in the templates are scaled proportionally or adjusted based on rules to obtain a multi-dimensional set of standard task coefficients tailored to the current real-time business requirements.

[0050] Each component in the multidimensional standard task coefficient set represents the number of times the corresponding standard computation task needs to be performed to complete the current real-time business requirements. For example, the standard business requirements set for industrial visual quality inspection is: image acquisition: 10 times, image enhancement: 50 times, defect segmentation: 100 times, feature classification: 20 times, result reporting: 5 times. The current business requires inspection of products with higher precision, so the number of defect segmentation operations needs to be increased to 200 times. Therefore, the output multidimensional standard task coefficient set is: [10, 15, 50, 200, 20].

[0051] Furthermore, based on the multidimensional standard task coefficient set and the standard business requirement set, the multidimensional computing power requirement of real-time business needs is determined. The multidimensional computing power requirement is a vector composed of quantitative indicators of various computing resources required to complete the current real-time business needs. The dimension of the vector corresponds to the resource type that the scheduling system focuses on.

[0052] Specifically, the system receives a multi-dimensional standard task coefficient set of length K and simultaneously reads the pre-stored standard task-resource consumption matrix from the standard business requirement set. Then, it multiplies the multi-dimensional standard task coefficient set with the standard task-resource consumption matrix to obtain a new M-dimensional row vector, which represents the multi-dimensional computing power requirement of the real-time business needs. For example, there are 3 types of standard tasks and 2 resource dimensions, the multi-dimensional standard task coefficient set is [a, b, c], and the first column of the consumption matrix is ​​[p1, p2, p3]. T The second column (memory bandwidth) is [q1, q2, q3]. TThen, standard business requirement 1 is the sum of the products of a and p1, p2, p3, and standard business requirement 1a is the sum of the products of q1, q2, q3. The resulting vector is the multidimensional computing power requirement.

[0053] Furthermore, based on the business process information and business type, the corresponding business constraint information is obtained, which includes at least the upper bound of latency, the lower bound of reliability, and the security level.

[0054] The upper bound of latency is the maximum allowed time interval from the initiation of a business request to the completion of the business execution; the lower bound of reliability is the minimum requirement for the probability of successful completion of the business within a specified time range, reflecting the business's tolerance for task failure, node failure, or network interruption; the security level is the level of business requirements in terms of data confidentiality, integrity, availability, and privacy protection. Security levels are usually divided into several levels, with different levels corresponding to different encryption algorithms, access control policies, and isolation requirements.

[0055] Specifically, the process begins by parsing the explicit constraint fields in the original business requirements; if they exist, they are read directly. If the original requirements do not explicitly provide certain constraints, default constraint values ​​for the identified business type are retrieved from the pre-configured policy library. The determined upper bound of latency, lower bound of reliability, security level, and other constraint information are then formatted and stored to form business constraint information. The security level can be mapped to integer levels, such as 1, 2, 3, from low to high, with larger numbers indicating higher security levels.

[0056] Finally, the vectorized multidimensional computing power demand and business constraint information are concatenated into real-time business components.

[0057] Specifically, the multi-dimensional computing power demand vector and business constraint information are fused. First, the multi-dimensional computing power demand is vectorized to obtain an M-dimensional floating-point vector. Then, the business constraint information is normalized. Finally, the computing power demand vector and the constraint scalar are concatenated in a predetermined order to form a vector, which is used as the real-time business component. For example, if the multi-dimensional computing power demand is [445, 109.15] and the business constraint information is [0.05, 0.9999, 3], then the real-time business component vector is [445, 109.15, 0.05, 0.9999, 3].

[0058] In step S20 of the method provided in this application embodiment, the pre-construction of the standard business requirement set includes: A predefined set of standard computing tasks is provided, which includes multiple benchmark computing tasks associated with parameterized computing features, and each benchmark computing task represents a type of computing mode. For each of the benchmark computing tasks, offline benchmark performance tests are performed on multiple computing nodes in the target computing power network to obtain the original performance characterization values ​​of the multiple computing nodes for each of the benchmark computing tasks. Using the statistical extreme values ​​of all the computing power nodes as a normalization benchmark, the original performance characterization values ​​are normalized to obtain the normalized performance characterization values ​​of each computing power node for each benchmark computing task. Establish a mapping relationship between business types and the standard computing task set. For each business type, determine the proportion coefficient of each benchmark computing task according to its typical execution process to form the standard business requirement set.

[0059] In this embodiment of the application, a standard computing task set is first predefined. The standard computing task set includes multiple benchmark computing tasks associated with parameterized computing characteristics, and each benchmark computing task represents a type of computing mode. The standard computing task set is a finite set of discrete tasks that constitutes an atomic unit or benchmark scale for measuring the computing power of all computing nodes and the resource requirements of all business. The parameterized computing characteristics are the parameters associated with the load characteristics of each benchmark computing task, in addition to its name and functional description.

[0060] A benchmark computing task is a computational load with defined input and output specifications, and has a clear reference implementation and standard test dataset; a computing mode is a common abstraction of the resource consumption characteristics of a class of computing tasks, and each benchmark computing task should represent a combination of one or more computing modes.

[0061] Specifically, the first step is to analyze the business domain targeted by the target computing network to identify common underlying computing patterns. Then, for each computing pattern, one or more specific benchmark computing tasks are defined and associated with complete parameterized computing characteristics, such as input data size, algorithm complexity, and memory access patterns.

[0062] All the established benchmark computation tasks together constitute the standard computation task set. For example, five benchmark computation tasks can be set as follows: T1: Image scaling - bilinear interpolation, representing a memory-intensive mode; T2: Face detection - MTCNN inference, representing an AI inference mode; T3: Point cloud registration - ICP algorithm, representing a floating-point intensive mode; T4: AES-256 encryption, representing an integer-intensive mode; and T5: Video encoding - H.264, representing a hybrid computation and memory access mode. Each task is associated with parameters such as input data size, algorithm complexity order, and memory access mode.

[0063] Secondly, for each benchmark computing task, offline benchmark performance tests are conducted on multiple computing nodes in the target computing power network to obtain the original performance characterization values ​​of multiple computing nodes for each benchmark computing task. The offline benchmark performance test is a test performed before the computing power network carries real-time services or during the maintenance window to measure its performance in executing a specific benchmark computing task. The original performance characterization values ​​are the raw performance metrics obtained through testing without any mathematical transformation.

[0064] Specifically, for each benchmark computing task in the standard computing task set, a standardized test program for the task is deployed and executed on each selected computing node in the target computing power network. The test program typically uses the same input data, the same algorithm implementation, and the same compilation optimization options to ensure that the test results reflect only the performance differences of the node hardware and the underlying software stack. Each test is repeated multiple times, for example, 100 times, and the average value is taken as the raw performance characterization value after the number of repetitions is completed.

[0065] Subsequently, the original values ​​are recorded in the performance matrix. The rows of the matrix correspond to computing nodes, the columns correspond to standard computing tasks, and the matrix elements are the original performance characterization values. For example, the target computing network contains three nodes: cloud node C, edge node E, and terminal node D. Offline tests are performed on the standard task sets T1 to T5. For T2: face detection, the measured original performance characterization values ​​are: node C 2000 times / second, node E 500 times / second, and node D 15 times / second. For T4: AES-256 encryption, the measured values ​​are: node C 5000 MB / s, node E 800 MB / s, and node D 20 MB / s. These are used as the original performance characterization values ​​and filled into the performance matrix.

[0066] Next, the original performance characterization values ​​are normalized using the statistical extreme values ​​of all computing nodes as the normalization benchmark. This yields the normalized performance characterization values ​​of each computing node for each benchmark computing task. Normalization is a process of mapping values ​​of different dimensions and magnitudes to a common, dimensionless numerical range through mathematical transformation. Commonly used normalization methods include min-max normalization and Z-score standardization.

[0067] Specifically, the maximum and minimum values ​​of the original performance characterization values ​​of the benchmark computation task are used as the normalized baseline values. Min-max normalization is applied to normalize the original performance characterization values, and the normalized value is calculated as (original value - minimum value) / (maximum value - minimum value). For some tasks, such as throughput, higher performance is better; for others, such as latency, lower performance is better. A value of 1 indicates that the node performs best on this task, and a value of 0 indicates the worst performance. By normalizing the performance characterization values, the absolute magnitude differences between nodes are eliminated, ultimately yielding the normalized performance characterization value for each node for each standard task.

[0068] For example, for T2: face detection, after normalization, node C=1.0, node E=0.244, and node D=0. For T5: video coding, assuming the original performance: C=1000 frames / second, E=300 frames / second, and D=10 frames / second, then after normalization, C=1.0, E=(300-10) / (1000-10)=290 / 990≈0.293, and D=0. The normalized value of node E is 0.244 on T2 and 0.293 on T5, indicating that video coding has a slightly better relative performance than face detection. The normalized value serves as the dimension of the initial feature vector of node E in the graph.

[0069] Finally, a mapping relationship is established between business types and standard computing task sets. For each business type, the proportion coefficient of each benchmark computing task is determined according to its typical execution process to form a standard business requirement set. The mapping relationship is a correspondence rule from business type to a set of quantified coefficients. The mapping relationship is usually stored in the form of a lookup table, configuration file or rule engine. The typical execution process is the sequence of standard computing steps that the business logic goes through under standardized input parameters for a certain business type.

[0070] Specifically, for each predefined business type, the proportion coefficient of each benchmark calculation task in its typical execution process is determined, and the proportion coefficient is associated with the business type and stored to form a record of the standard business requirement set. This process is repeated for all business types to finally form a complete standard business requirement set.

[0071] The typical execution process can be obtained by analyzing the application architecture of the business, expert knowledge, or by statistical analysis of a large number of actual business trajectories. That is, the proportion coefficient of each benchmark computing task is obtained by the following methods: through expert experience, domain experts manually specify according to business logic; through static code analysis, analyze the executable files or intermediate representations of the business application and count the number of calls to various underlying operations; through runtime profiling, run typical business in a controlled environment, record the execution trajectory through performance profiling tools, and count the frequency of various computing operations.

[0072] For example, consider the typical workflow of facial recognition: image acquisition → face detection → key point localization → feature extraction → feature comparison. Face detection can be mapped to T2, feature extraction to a variant of T2, and feature comparison to T1. Assuming a baseline configuration, completing one facial recognition task requires 1 face detection, 1 feature extraction, and 1000 feature comparisons, the proportions are: T1: 1000, T2: 1.5, and other tasks are 0.

[0073] In this embodiment, a unified metric is established for heterogeneous computing power nodes by pre-constructing a standard business requirement set. Then, the business type of real-time business requirements is identified, business process information is invoked, and multi-dimensional computing power requirements are determined. At the same time, business constraint information is extracted and finally vectorized and concatenated into real-time business components, reducing the computational latency of real-time scheduling, avoiding resource mismatch caused by matching a single indicator, and improving maintainability.

[0074] S30: Obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, combine the real-time service components with the heterogeneous computing resource characteristic map, perform optimal route solving based on multi-objective optimization, and obtain the optimal computing route; In this embodiment, the weight coefficient of the multidimensional scheduling strategy is the degree of preference for a specific scheduling optimization objective corresponding to each component; multi-objective optimization is to find a set of Pareto optimal solutions when there are multiple conflicting objective functions; optimal routing solution is, in the context of computing power network, routing is no longer just the transmission path of network data packets, but also the execution path of computing tasks; alternative routing schemes are the set of candidate computing paths that meet the basic connectivity requirements generated by the algorithm during the solution process.

[0075] Specifically, firstly, based on the current application scenario, the weight coefficients of the multi-dimensional scheduling strategy are obtained. Then, combining real-time business components and heterogeneous computing resource characteristic maps, a multi-objective optimization problem is constructed. Subsequently, the lower bound of the search space is limited to the edge side, generating multiple alternative routing schemes. For each alternative scheme, the cost value on each optimization objective is calculated using node embedding vectors and relational edge information, and the total task cost value is obtained by weighted summation. Finally, the alternative routing scheme with the lowest total task cost value is selected as the optimal computational route.

[0076] Step S30 in the method provided in this application embodiment includes: Based on the prior multidimensional evaluation rules of the multidimensional scheduling strategy and the target scenario, the weight coefficient of the multidimensional scheduling strategy is determined. The multidimensional scheduling strategy includes cost-aware scheduling strategy, load-aware scheduling strategy, energy-efficiency-aware scheduling strategy and service level agreement-aware scheduling strategy. Combining the multi-dimensional scheduling strategy, the weight coefficients of the multi-dimensional scheduling strategy, and the real-time service components, a multi-objective optimization task is constructed, wherein the objective function of the multi-objective optimization task is the weighted sum of the weight coefficients of the multi-dimensional scheduling strategy and the sub-objective function values ​​corresponding to each scheduling strategy; The search lower limit of the heterogeneous computing resource characteristic map is limited to the edge side, multiple alternative routing schemes are generated, and the task cost value of the multiple alternative routing schemes for the multi-objective optimization solution task is calculated and obtained. The optimal computation route is determined based on the task cost.

[0077] In this embodiment of the application, the weight coefficient of the multi-dimensional scheduling strategy is first determined based on the prior multi-dimensional evaluation rules of the multi-dimensional scheduling strategy and the target scenario. The multi-dimensional scheduling strategy includes cost-aware scheduling strategy, load-aware scheduling strategy, energy-efficiency-aware scheduling strategy and service level agreement-aware scheduling strategy.

[0078] Specifically, multi-dimensional scheduling strategies involve considering multiple optimization directions or objectives simultaneously during the scheduling decision-making process. Each strategy corresponds to a quantifiable evaluation dimension. Cost-aware scheduling strategies prioritize paths with the lowest computational or energy costs; load-aware scheduling strategies prioritize nodes with the lightest current load and the most abundant remaining computing power; energy-efficiency-aware scheduling strategies prioritize nodes with the lowest energy consumption per unit task, suitable for green computing scenarios; and service level agreement-aware scheduling strategies prioritize ensuring the service quality commitments of the business.

[0079] The target scenario refers to the specific application environment or operational stage of the computing network. Different target scenarios have different degrees of emphasis on the above scheduling strategies; the prior multidimensional evaluation rules are pre-set rules or formulas used to determine the importance of each scheduling strategy; the multidimensional scheduling strategy weight coefficients are the weights of cost perception, load perception, energy efficiency perception, and service level agreement perception strategies, used to reflect the operator's preferences in the trade-off between multiple objectives.

[0080] First, the current target scenario identifier is obtained. Then, pre-stored prior multi-dimensional evaluation rules are invoked. For example, the rules could be: if the scenario is cost-priority, the weight coefficients are set to [0.7, 0.1, 0.1, 0.1]; if the scenario is performance-priority, they are set to [0.1, 0.2, 0.1, 0.6]; if the scenario is load-balancing, they are set to [0.1, 0.7, 0.1, 0.1]; if the scenario is green and energy-saving, they are set to [0.2, 0.1, 0.6, 0.1]. Subsequently, the corresponding weight coefficients are read according to the current target scenario, and a four-dimensional weight vector is output. The output multi-dimensional scheduling strategy weight coefficients will directly participate in the weighted summation calculation of each objective function in the subsequent multi-objective optimization solution task.

[0081] Secondly, by combining multi-dimensional scheduling strategies, multi-dimensional scheduling strategy weight coefficients, and real-time service components, a multi-objective optimization task is constructed. The multi-objective optimization task is a well-defined optimization problem, and its objective function is the weighted sum of the multi-dimensional scheduling strategy weight coefficients and the sub-objective function values ​​corresponding to each scheduling strategy.

[0082] Specifically, based on the cost function of the multi-dimensional scheduling strategy, the output weight coefficients, and the real-time service components, a multi-objective optimization task is constructed. In the multi-dimensional scheduling strategy, each strategy corresponds to a function factor, the weights correspond to the weighted sum weights, the decision variable is the scheduling path, and the constraints include at least resource capacity constraints and service quality constraints. The constructed optimization task is to minimize the cost function value while satisfying the constraints.

[0083] The multi-objective optimization task is defined as follows: ω1 × cost requirement + ω2 × load requirement + ω3 × energy efficiency requirement + ω4 × service level agreement requirement, where ω1, ω2, ω3, and ω4 are the weight coefficients of the cost-aware scheduling strategy, load-aware scheduling strategy, energy efficiency-aware scheduling strategy, and service level agreement-aware scheduling strategy, respectively. The constraints include the available computing power, prediction reliability, and node security level of each node on the route.

[0084] Furthermore, the search lower bound of the heterogeneous computing resource characteristic map is limited to the edge side. Multiple alternative routing schemes are generated, and the task cost value of multiple alternative routing schemes for the multi-objective optimization solution is calculated. Here, the search lower bound is the lowest level allowed by the routing scheme in the topology; the alternative routing scheme is the set of all possible paths from the source node to the target node in the heterogeneous computing resource characteristic map; and the task cost value is the value calculated by substituting each alternative routing scheme into the objective function.

[0085] Finally, the optimal computational route is determined based on the task cost value, where the task cost value is the corresponding value of each alternative route under the multi-objective weighted evaluation. The smaller the value, the better the route is under the multi-objective weighted evaluation.

[0086] Specifically, the system receives a list of candidate routing schemes and their corresponding task costs, sorting them from lowest to highest cost. Then, it selects the scheme with the lowest cost as the optimal computation route. For example, if the three candidate schemes have costs of 5.2, 6.1, and 8.7 respectively, the scheme with a cost of 5.2 is selected as the optimal computation route. The system then outputs complete path information, including the identifiers of the nodes traversed in sequence, the computational task share each node needs to undertake, and the network link information along the path. This information is then passed to subsequent edge-to-end task allocation.

[0087] In this embodiment, weight coefficients are determined by prior multidimensional evaluation rules based on multidimensional scheduling strategies and target scene calibration, and a multi-objective optimization task is constructed to take into account multiple conflicting objectives and avoid the one-sidedness of single-objective optimization. The search lower limit is limited to generating alternative routing schemes on the edge side and calculating the task cost value. Finally, the optimal computation route is determined based on the cost value to avoid scheduling the task to overloaded or faulty nodes. At the same time, the comparison of task cost values ​​provides an explanatory basis for route selection.

[0088] S40: Based on the optimal computing route and the heterogeneous computing resource characteristic map, construct a multi-dimensional computing power spectrum for the subnet, and combine it with the real-time service components to allocate edge-end tasks; In this embodiment, the subnet multidimensional computing power spectrum is a local resource view after dimensionality reduction, where each point represents the multidimensional computing power capability of an end node; the subnet multidimensional computing power spectrum space is a vector space spanned by the multidimensional computing power spectrum; the dimensional demand radius is the demand region that needs to be satisfied in the subnet multidimensional computing power spectrum space, which is the vector of multidimensional computing power demand in the real-time service components; the demand value in each dimension is the demand radius in that dimension; the greedy algorithm is an algorithm that takes the best or optimal choice in the current state at each step, hoping to lead to the global best or optimal result.

[0089] Specifically, the algorithm first updates the heterogeneous computing resource characteristic map. Then, using the identified edge nodes as indices, it extracts the feature embedding vectors of all end-side nodes belonging to the edge nodes, constructing a local subnet multi-dimensional computing power spectrum. Next, using the business constraint information in the real-time business components as a hard screening condition, it removes all end-side nodes that do not meet the constraints. The multi-dimensional computing power demand in the real-time business components is considered as the multi-dimensional demand points to be covered, and its demand radius in each dimension of the subnet multi-dimensional computing power spectrum space is calculated. The algorithm first determines whether there exists an end-side node whose computing power values ​​in all dimensions are greater than or equal to the corresponding dimension demand radius. If so, a greedy strategy is directly adopted to select the node closest to the demand point as the target task node, completing the edge-end task allocation.

[0090] Step S40 in the method provided in this application embodiment includes: Update the heterogeneous computing resource characteristic map, and extract the feature embedding vectors of the path nodes using the optimal computing route as an index; Based on multiple feature embedding vectors, and using multiple computing power dimensions as coordinates, the multiple feature embedding vectors are mapped to the corresponding multi-dimensional space to construct the subnet multi-dimensional computing power spectrum and form the subnet multi-dimensional computing power spectrum space. Using the business constraint information in the real-time business components as hard constraints, unqualified end-side computing power nodes in the subnet multi-dimensional computing power spectrum space are screened out; The multidimensional computing power demand in the real-time service components is decomposed into the dimensional demand radius of each dimension in the multidimensional computing power spectrum space of the subnet. Determine whether there are edge computing power nodes above the required radius of multiple dimensions. If so, select the edge computing power node as the target task node using a greedy algorithm and output the edge-end task allocation result accordingly.

[0091] In this embodiment, the heterogeneous computing resource characteristic map is first updated, and the feature embedding vector of the path node is extracted using the optimal computing route as the index. The update is triggered again before the edge-end task allocation begins.

[0092] Specifically, a graph neural network model is invoked, and node state parameters collected by the cloud-edge-device perception layer are used to perform an incremental embedding update of the heterogeneous computing resource characteristic map. This ensures that the feature embedding vectors in the map reflect the actual load, available computing power, and network status of each node at the current moment. Then, the optimal computing route is obtained, each node identifier is traversed, and the corresponding node object is retrieved in the updated map, and its current feature embedding vector is read. For example, if the optimal computing route is [device node D, edge node E1, cloud node C, edge node E1], the feature embedding vector of edge node E1 (dimension 16) and the feature embedding vector of cloud node C are extracted, etc. For nodes that appear repeatedly, they can be extracted only once or processed separately according to the context. The extracted feature embedding vectors are used in the next step to construct the subnet multi-dimensional computing power spectrum.

[0093] Secondly, based on multiple feature embedding vectors and using the values ​​of multiple computing power dimensions as coordinates, the multiple feature embedding vectors are mapped to the corresponding multi-dimensional space to construct the subnet multi-dimensional computing power spectrum and form the subnet multi-dimensional computing power spectrum space.

[0094] Specifically, first, the last node of the optimal computation route is determined. Then, the feature embedding vectors of all child nodes of this edge node are obtained from the heterogeneous computing resource characteristic map. Since one dimension of the feature embedding vector encodes computing power information and another part encodes state or topology information, the scheduling-related M computing power dimension is extracted from the embedding vector of each edge node according to a preset mapping rule. For example, M=3, corresponding to AI computing power, memory bandwidth, and storage capacity. The values ​​of the M dimensions are used as the coordinates of the edge node in the multi-dimensional computing power spectrum space. The coordinates of all edge nodes together constitute the subnet multi-dimensional computing power spectrum. The three-dimensional space in which the node is located is the subnet multi-dimensional computing power spectrum space.

[0095] Secondly, using the business constraint information in the real-time business components as hard constraints, unqualified edge computing power nodes in the subnet's multi-dimensional computing power spectrum space are filtered out. Hard constraints are restrictions that must be unconditionally met during task allocation; violating any hard constraint will invalidate the allocation result. Filtering involves checking all edge nodes to see if they meet all hard constraints. Nodes that do not meet any hard constraint are removed from the candidate node set and do not participate in subsequent matching. Unqualified edge computing power nodes are edge nodes that do not meet at least one requirement in the business constraint information.

[0096] Specifically, the edge nodes within the subnet are filtered. First, the service constraint information in the real-time service components is parsed to obtain three key values: upper bound of latency, lower bound of reliability, and security level. Then, for each edge node in the multi-dimensional computing power spectrum space of the subnet, latency checks, reliability checks, and security checks are performed. For latency checks, the network latency from the edge node to its home edge node is obtained from the graph. This can be checked using the latency component in the edge weights; if the latency is greater than the upper bound, it is marked as unqualified. For reliability checks, the historical reliability records of the edge node are obtained from the graph. If the obtained reliability is less than the lower bound, it is marked as unqualified. For security level checks, the security level identifier of the edge node is obtained from the graph. If the level is less than the corresponding security level, it is marked as unqualified.

[0097] If any check fails, the corresponding node is eliminated. This process is repeated for all nodes, filtering out unqualified nodes. The remaining nodes form a candidate node set. If the candidate set is empty, single-node allocation is not possible, triggering a multi-node allocation or task splitting process. This hard-constraint filtering ensures that all endpoint nodes entering the subsequent matching stage have the basic qualifications to undertake the task.

[0098] For example, the business constraints in the real-time business component are: upper bound of latency 5ms, lower bound of reliability 99.9%, and security level 2. Assume there are three end-side nodes in the subnet's multi-dimensional computing power spectrum space: D1: latency 3ms, reliability 99.95%, security level 3, deemed qualified; D2: latency 8ms, reliability 99.9%, security level 2, deemed latency exceeds the limit and eliminated; D3: latency 2ms, reliability 99.5%, security level 2, deemed insufficient reliability and eliminated. The final candidate node set includes D1.

[0099] Furthermore, the multidimensional computing power demand in the real-time business components is decomposed into the dimensional demand radius of each dimension in the multidimensional computing power spectrum space of the subnet, where the multidimensional computing power demand represents the absolute quantity of various computing resources required to complete the business.

[0100] Specifically, firstly, the maximum and minimum values ​​of all end-side nodes in the subnet's multi-dimensional computing power spectrum space are obtained in each computing power dimension. Then, normalization is performed on each dimension of the multi-dimensional computing power demand in the real-time business components. The normalization formula for each dimension is: (absolute demand of computing power demand in the corresponding dimension - minimum actual capability of nodes in the subnet) / (maximum actual capability of nodes in the subnet - minimum actual capability of nodes in the subnet).

[0101] If the absolute computing power requirement is less than the minimum actual capacity, the calculated value may be negative and can be truncated to 0, indicating that the requirement for that dimension is extremely low and can be met by any node. If the absolute computing power requirement is greater than the maximum actual capacity, the calculated value is greater than 1, indicating that a single node cannot meet the requirement for that dimension, triggering multi-node allocation or task splitting. The normalized vector is the dimensional requirement radius for each dimension. In a multi-dimensional space, for each dimension's requirement radius, the coordinates of the endpoint node are considered a qualified single-node candidate only if the coordinates of each dimension are greater than or equal to the corresponding requirement radius.

[0102] For example, in dimension 1 of the subnet's multidimensional computing power spectrum space, the node capability range is [0, 100]. The multidimensional computing power requirement for real-time services is 45 TOPS. After normalization, the radius of a certain dimension's requirement is (45-0) / (100-0) = 0.45.

[0103] Finally, it is determined whether there are edge computing power nodes above the demand radius in multiple dimensions. If so, a greedy algorithm is used to select the edge computing power node as the target task node, and the corresponding edge-to-edge task allocation result is output. The greedy algorithm selects the node with the largest surplus exceeding the demand radius in all dimensions from among multiple candidate nodes that meet the conditions, or selects the node with the closest Euclidean distance to the demand point. The greedy algorithm does not guarantee global optimality, but it is computationally efficient.

[0104] Specifically, firstly, the candidate end-side nodes obtained after screening are compared one by one based on the dimensional requirement radius vector. For each candidate node, it is checked whether its coordinates in each computing power dimension satisfy the condition that the coordinates are greater than the dimensional requirement radius vector. If at least one node satisfies this condition, the single-node allocation branch is entered. A greedy algorithm is used to select an optimal node from the nodes that meet the condition.

[0105] Greedy strategies can be pre-configured. For example, one approach is to select the node with the largest overall capacity surplus. This involves calculating the sum of the differences between each node's coordinates in each dimension and the demand radius, and selecting the node with the largest surplus, indicating that the node has more leeway to cope with future load fluctuations. Alternatively, the approach can be to select the node closest to the demand. This involves calculating the node with the closest Euclidean distance to the demand point to avoid resource waste. The Euclidean distance is the square root of the sum of the squares of the differences in each dimension. After calculating the Euclidean distance for all nodes, the node with the smallest Euclidean distance is selected. After selecting a node, the edge-to-end task allocation result is generated and output for collaborative scheduling. If no single node meets the conditions, the process proceeds to multi-node allocation or task splitting.

[0106] The cloud-edge-device collaborative scheduling method for computing power networks provided in this application embodiment further includes: If there are no edge computing nodes above the required radii of multiple dimensions, then a set of edge computing nodes is selected in the multidimensional computing power spectrum space of the subnet based on a greedy algorithm. Determine whether the sum of the edge computing power node set satisfies the required radius of multiple dimensions. If it does, output the edge computing power node set as the target task node set and output the edge-end task allocation result accordingly. Otherwise, tasks are split based on the real-time business components, and edge-end task allocation is performed on the multiple split tasks in the task splitting results.

[0107] In this embodiment of the application, if there are no end-side computing power nodes above the multi-dimensional demand radius, then a set of end-side computing power nodes is selected in the multi-dimensional computing power spectrum space of the subnet based on a greedy algorithm.

[0108] Specifically, when it is determined that there are no edge computing power nodes, the process enters the multi-node collaborative allocation branch. First, all candidate edge nodes in the current subnet's multi-dimensional computing power spectrum space, after being filtered by hard constraints, are obtained. Then, the target node set is initialized, and the remaining demand vector is initialized to equal the original multi-dimensional computing power demand. An iterative greedy algorithm is used: in each round, all unselected candidate nodes are traversed, and for each node, its contribution to the current remaining demand in each dimension is calculated.

[0109] The greedy strategy selects the node with the largest surplus, maximizing the total amount of remaining demand filled. This node is then added to the initial target node set, and the remaining demand is updated. This process is repeated until all dimensions of the remaining demand are zero or no node can be found to fill any remaining demand. The final output is the selected set of edge computing nodes.

[0110] For example, suppose an edge node has three endpoint nodes: D1 (0.4, 0.3), D2 (0.3, 0.6), and D3 (0.5, 0.2). The business demand radius is (0.6, 0.5). Since no single node can satisfy the demand, a greedy algorithm is used, with the initial remaining demand (0.6, 0.5). In the first round, calculate the contribution of each node: D1 contribution = 0.4 + 0.3 = 0.7; D2 contribution = 0.3 + 0.5 = 0.8; D3 contribution = 0.5 + 0.2 = 0.7. Select the node with the largest contribution, D2. At the same time, update the remaining demand (0.3, 0.0). In the second round, with the remaining demand (0.3, 0.0), the candidate nodes D1 contribution = 0.3 + 0 = 0.3 and D3 contribution = 0.3 + 0 = 0.3 are selected, and the remaining demand (0.0, 0.0) is updated. The final selected set is [D2, D1].

[0111] Furthermore, it is determined whether the sum of the edge computing power node sets meets the multi-dimensional requirement radius. If it does, the edge computing power node set is output as the target task node set, and the corresponding edge-end task allocation result is output. The target task node set is the same as the edge computing power node set, and it is confirmed that it meets the requirements after summation.

[0112] Specifically, the validity of the edge node set obtained through the greedy selection is verified by calculating the total capacity of the edge node set across each computing power dimension. Then, the total capacity across each computing power dimension is compared with the required radius of that dimension. If the total capacity across all dimensions is greater than or equal to the required radius of that dimension, the selection is successful. At this point, the edge node set is output as the target task node set, and the corresponding edge-to-edge task allocation result is generated.

[0113] The output edge-to-end task allocation results should include: a list of nodes and a description of the subtasks undertaken by each node, the input data shards that each node needs to receive, the expected execution sequence, and the result aggregation method.

[0114] If any dimension is not satisfied, the judgment fails, and the task splitting process begins. Since the goal of the greedy algorithm is to make the set satisfy the requirements as much as possible, but the node capabilities are discrete and indivisible, there may be situations where the greedy algorithm cannot find a combination that completely satisfies the requirements, or the total capability of all nodes in the subnet is insufficient to meet the requirements.

[0115] For example, suppose the greedy algorithm selects the edge node set [D2, D1]. D2 has coordinates (0.3, 0.6), D1 has coordinates (0.4, 0.3), and their sum is (0.7, 0.9). The required radius is (0.6, 0.5). 0.7 ≥ 0.6 and 0.9 ≥ 0.5, which satisfies the condition. Therefore, the output target task node set is [D1, D2], and the allocation result is generated: D1 is responsible for feature extraction, D2 is responsible for classification, the result of D1 needs to be sent to D2, and finally D2 returns the result to the edge node.

[0116] Otherwise, tasks are split based on real-time business components, and edge-to-end task allocation is performed on the multiple split tasks in the result. Task splitting involves dividing complex, resource-intensive real-time business requirements into multiple logically independent subtasks with smaller resource requirements according to their business process information or task decomposition templates in a standard business requirement set. The split subtasks can be executed serially or in parallel, and each subtask can be independently scheduled to different computing nodes for execution.

[0117] Specifically, if the summation fails to determine the result, it indicates that all edge nodes within the current edge subnet cannot jointly undertake the original service. Therefore, the service process information in the real-time service components is obtained, and based on this process, the original task is broken down into multiple sub-tasks along the functional boundaries. For example, image acquisition and preprocessing are merged into one sub-task, and feature extraction and classification are merged into another sub-task.

[0118] Each split task is associated with a new real-time business component. The multi-dimensional computing power requirement is the original requirement allocated to that sub-task, and the corresponding business constraints also need to be adjusted accordingly. For example, if the original total latency was 100ms, after splitting, each sub-task might be allocated 30ms, 40ms, or 30ms. Then, for each split task, the edge-to-end task allocation process is re-executed, namely, constructing the subnet multi-dimensional computing power spectrum, hard constraint filtering, requirement decomposition, and single-node / multi-node allocation. Finally, the allocation results of all split tasks are aggregated to form a multi-node collaborative scheduling scheme.

[0119] In this embodiment, by updating the heterogeneous computing resource characteristic map and extracting feature embedding vectors using the optimal computing route as an index, a multi-dimensional computing power spectrum of the subnet is constructed to form a spectrum space, avoiding resource mismatch due to single-dimensional matching; unqualified end-side nodes are screened out using business constraint information as hard constraints, and it is determined whether there is a single node that meets the requirements, ensuring that the end-side nodes meet the key requirements and guaranteeing the quality of business services; the multi-dimensional computing power requirement is decomposed into dimensional requirement radii, and an end-side node set is constructed. The end-side node set is selected by a greedy algorithm and summed for discrimination, realizing the multi-node aggregation of end-side computing power, breaking through the resource boundaries of a single subnet, while maintaining the correctness of data dependency relationships.

[0120] S50: Integrate the optimal computational route with the edge-end task allocation results to perform collaborative scheduling of real-time business requirements.

[0121] In this embodiment, collaborative scheduling is the fused execution plan that coordinates the cloud, edge, and terminal computing power nodes to complete the unloading, transmission, computation, and result feedback of computing tasks; collaborative scheduling of real-time business requirements is to deploy and execute the initially input real-time business requirements on the entire computing power network according to the scheduling plan.

[0122] Specifically, the optimal computation route and edge-to-device task allocation results are fused to obtain a complete scheduling path. Then, according to the fused path, a specific instruction sequence is generated and distributed to the relevant cloud, edge, and device nodes. Each node works collaboratively according to the instructions: device nodes are responsible for data collection and lightweight computing, edge nodes are responsible for routing and forwarding and task decomposition, and the cloud center is responsible for heavy-duty computing. Ultimately, under the premise of meeting business constraints, the collaborative processing of the entire real-time business is completed.

[0123] In step S50 of the method provided in this application embodiment, the optimal computational route and the edge-end task allocation results are fused to perform collaborative scheduling of real-time service requirements. Afterwards, the method further includes: Multiple edge computing nodes perform local computations based on the edge-to-edge task allocation results and periodically report their execution status to the corresponding edge computing nodes. The execution status includes at least the task completion progress and resource utilization rate. The edge computing nodes update the subnet's multidimensional computing power spectrum based on the reported status. When the state change rate of an edge computing node exceeds a preset rematch threshold, real-time rematching based on spectral position sliding is triggered, including: If the load increases, it will first expand to the neighborhood of the current spectral position. If the neighborhood is insufficient, it will expand to higher-level spectral positions along the spectral axis. If the load decreases, higher-level spectral positions are released first, while lower-level spectral positions contract along the spectral axis. The real-time rematching results are merged with the status reporting results and aggregated to the cloud computing nodes to update the heterogeneous computing resource characteristic map.

[0124] In this embodiment of the application, multiple edge computing nodes first perform local computation based on the edge-end task allocation result, and periodically report the execution status to the corresponding edge computing nodes. The execution status includes at least the task completion progress and resource utilization rate. Periodic reporting means that the edge nodes send the execution status to the edge computing nodes to which they belong at fixed time intervals. The execution status is a quantitative indicator of the current task execution status of the edge nodes.

[0125] Specifically, after completing the coordinated scheduling, the specific computation task instructions and input data are sent to the designated edge nodes through the optimal computation route. Each edge node assigned a task begins executing local computation. Simultaneously, the edge node starts a periodic reporting timer. Every preset period, such as 200 milliseconds, the node collects the current task completion progress, such as 150 frames processed (15% of the total 1000 frames), and resource utilization, such as CPU utilization of 75%, memory utilization of 60%, and NPU utilization of 90%. This data is then integrated into a status report message and sent to the corresponding edge computing node through its communication interface. Upon receiving the status report, the edge node stores it in its local cache for subsequent subnet multi-dimensional computing power spectrum updates and re-matching decisions.

[0126] Secondly, the edge computing nodes update the subnet's multidimensional computing power spectrum based on the status reporting results.

[0127] Specifically, whenever an edge computing node receives a status report from one of its subordinate end-side nodes, it parses the corresponding node's resource utilization. Then, the edge node reads the original normalized performance vector of that end-side node from its local cache. For each computing power dimension, the edge node calculates the updated available computing power coordinates. It replaces the old coordinates of that node in the subnet's multi-dimensional computing power spectrum with the calculated available computing power coordinates. After the update is complete, the edge node stores the new subnet multi-dimensional computing power spectrum locally for allocation decisions of subsequently arriving tasks.

[0128] Furthermore, when the state change rate of an edge computing node exceeds a preset rematch threshold, real-time rematching based on spectral position sliding is triggered. This includes: if the load increases, prioritizing expansion to the neighborhood of the current spectral position; if the neighborhood is insufficient, expanding to higher-level spectral positions along the spectral axis; if the load decreases, prioritizing the release of higher-level spectral positions and shrinking to lower-level spectral positions along the spectral axis. Here, the state change rate is the rate of change of the resource occupancy rate or available computing power of the edge node over time. A high change rate indicates a sharp increase or decrease in node load. The preset rematch threshold is a pre-configured numerical threshold used to determine whether the state change is drastic enough to require task reallocation.

[0129] Specifically, after each edge node receives a status report from the end-side node and updates the subnet's multidimensional computing power spectrum, it calculates the rate of change of the node's current report compared to the previous report, and then sets a preset rematch threshold. If the threshold is too low, it will lead to frequent rematches and increase the system burden; if the threshold is too high, it may miss key changes. Therefore, an appropriate value needs to be set to determine whether to trigger a rematch.

[0130] If the rate of change of state exceeds the preset rematch threshold, a rematch is triggered. The edge node first identifies the task currently being executed on the node. For each affected incomplete task, the edge node obtains the real-time business components of the task. If the load increases, it first expands to the neighborhood of the current spectral position, and performs incremental adjustments in the updated subnet multidimensional computing power spectrum space based on the task's required radius, re-executes the edge-end task allocation, and searches for other end-side nodes that meet the single-node or multi-node conditions.

[0131] If a suitable replacement node is found, the edge node will generate a task migration instruction. If no replacement node is found, or the neighborhood is insufficient, the task will be expanded along the spectral axis to a higher-level spectral position, either reverting to execution on the edge node itself or reporting it to the cloud for processing, thus gaining greater computing power support. For example, edge node E1 has a subordinate end-side node D1. Its last reported value was 0.8, with an interval of 200ms, resulting in a state change rate of |0.18 - 0.8| / 0.2 = 3.1 / s. The preset rematch threshold is 1.0 / s, triggering a rematch.

[0132] If the load decreases, higher-level spectral bits are released first, and lower-level spectral bits are contracted along the spectral axis. In other words, the task is readjusted to run in relatively basic or lower-priority spectral bits, thereby achieving dynamic saving and reasonable distribution of resources while ensuring task execution.

[0133] Finally, the real-time rematch results and status reporting results are merged and aggregated to the cloud computing nodes to update the heterogeneous computing resource characteristic map.

[0134] Specifically, after real-time rematching is completed, the results are fused with periodically accumulated status reporting data. Rematching events are marked as high-priority data and timestamped; status data from multiple edge nodes are sorted and compressed; subsequently, the fused data packet is sent to the cloud computing nodes. The cloud computing nodes parse the latest status parameters of each edge node. Then, using dynamic parameters, the feature embedding vectors of the corresponding nodes in the heterogeneous computing resource characteristic map are updated.

[0135] Nodes whose state changes exceed a threshold and their one-hop neighbors undergo local recalculation instead of a full graph update to reduce computational overhead. The updated graph is persistently stored and used for subsequent real-time service routing solutions.

[0136] In this embodiment, multiple edge nodes perform calculations based on the allocation results and report the execution status to the edge side, providing a data foundation for dynamic adjustment and realizing real-time observability of task execution. The edge side updates the subnet's multi-dimensional computing power spectrum based on the reported results, and subsequent new task allocation can avoid selecting overloaded nodes. When the state change rate of an edge node exceeds a preset rematch threshold, real-time rematching based on spectrum position sliding is triggered. Finally, the rematching results and the status reporting results are merged and summarized in the cloud to update the heterogeneous computing resource characteristic map, avoiding decision-making bias caused by outdated information.

[0137] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the target computing network is first subjected to hybrid sensing acquisition through the cloud-edge-device generalized sensing layer, and a graph structure is constructed based on the topological relationship. Then, the graph neural network is embedded and updated in combination with the node state parameters to achieve negative correlation adaptive adjustment between the acquisition cycle and the historical state change rate under the edge side, providing a unified and computable feature input basis for subsequent multi-objective optimization routing and task allocation.

[0138] Secondly, by pre-constructing a standard set of business requirements, a unified metric is established for heterogeneous computing power nodes. Then, the business type of real-time business requirements is identified, business process information is invoked, and multi-dimensional computing power requirements are determined. At the same time, business constraint information is extracted and finally vectorized and concatenated into real-time business components, reducing the computational latency of real-time scheduling, avoiding resource mismatch caused by matching a single indicator, and improving maintainability.

[0139] Furthermore, by determining the weight coefficients based on the multi-dimensional scheduling strategy and the prior multi-dimensional evaluation rules calibrated by the target scenario, a multi-objective optimization task is constructed to take into account multiple conflicting objectives and avoid the one-sidedness of single-objective optimization. The search lower limit is limited to generating alternative routing schemes on the edge side and calculating the task cost value. Finally, the optimal computation route is determined based on the cost value to avoid scheduling the task to overloaded or faulty nodes. At the same time, the comparison of task cost values ​​provides an explanatory basis for route selection.

[0140] Furthermore, by updating the heterogeneous computing resource characteristic map and extracting feature embedding vectors using the optimal computing route as an index, a multi-dimensional computing power spectrum of the subnet is constructed to form a spectral space, avoiding resource mismatch due to single-dimensional matching. Unqualified end-side nodes are screened out using business constraint information as hard constraints, and it is determined whether there is a single node that meets the requirements, ensuring that the end-side nodes meet the key requirements and guaranteeing the quality of business services. The multi-dimensional computing power requirement is decomposed into dimensional requirement radii, and an end-side node set is constructed. The end-side node set is selected through a greedy algorithm and summed for discrimination, realizing the multi-node aggregation of end-side computing power, breaking through the resource boundaries of a single subnet, while maintaining the correctness of data dependencies.

[0141] Ultimately, multiple edge nodes perform calculations based on the allocation results and report the execution status to the edge side, providing a data foundation for dynamic adjustments and achieving real-time observability of task execution. The edge side updates the subnet's multi-dimensional computing power spectrum based on the reported results, which can avoid selecting overloaded nodes in subsequent new task allocations. When the state change rate of an edge node exceeds a preset rematch threshold, real-time rematching based on spectrum position sliding is triggered. Finally, the rematching results and the status reporting results are merged and summarized in the cloud to update the heterogeneous computing resource characteristic map, avoiding decision-making biases caused by outdated information.

[0142] Example 2, as Figure 2 As shown, based on the same inventive concept as the cloud-edge-device collaborative scheduling method for computing power networks provided in Embodiment 1, this embodiment of the invention also provides a cloud-edge-device collaborative scheduling system for computing power networks, the system comprising: The embedding update module 11 is used to initialize and establish a heterogeneous computing resource characteristic map based on the topological relationship of the target computing power network, and to perform embedding update in combination with the node state parameters obtained by the cloud-edge-end general perception layer. The business requirement decomposition module 12 is used to obtain real-time business requirements and, in combination with a pre-built standard business requirement set, to perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components. The optimal route solving module 13 is used to obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, and combine the real-time service components and the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization to obtain the optimal computing route. The task allocation module 14 is used to construct a multi-dimensional computing power spectrum of the subnet based on the optimal computing route and the heterogeneous computing resource characteristic map, and to perform edge-end task allocation in combination with the real-time service components. The business demand scheduling module 15 is used to integrate the optimal computing route and the edge-end task allocation results to perform real-time collaborative scheduling of business demands.

[0143] In one embodiment, the embedded update module 11 is used for: The cloud-edge-device generalized perception layer is used to perform hybrid perception and acquisition on the target computing network to obtain node status parameters. Based on the aforementioned topological relationship, a graph structure is constructed using the relative performance values ​​of computing power nodes and computing power nodes facing multiple preset standard computing tasks as graph nodes, and the network distance between computing power nodes as relation edges. Based on the node state parameters, the graph nodes and the relation edges are embedded and updated in the graph structure to generate feature embedding vectors of multiple graph nodes, thereby obtaining the heterogeneous computing resource characteristic map. The feature embedding vector is defined as the embedding vector of itself updated by a nonlinear transformation function after weighting the neighbor information of neighboring nodes with edge weights.

[0144] Specifically, the cloud-edge-device generalized sensing layer performs hybrid sensing and acquisition on the target computing network to obtain node state parameters, including: The cloud-edge-device pervasive perception layer includes a cloud center perception agent, an edge-side perception agent, and a device-side perception agent. The hybrid perception acquisition includes event-driven acquisition and adaptive periodic acquisition. The cloud center perception agent and the edge side perception agent perform adaptive periodic data collection, wherein the collection period of the adaptive periodic data collection is negatively correlated with the historical state change rate under the edge side computing power node. The edge-side sensing agent and the endpoint-side sensing agent synchronously perform the adaptive periodic acquisition and the event-driven acquisition.

[0145] In one embodiment, the business requirement decomposition module 12 is used for: Identify the business type of the real-time business requirement, and invoke business process information based on the business type; By combining the business process information with the standard business requirement set, the real-time business requirements are decomposed into standardized requirements to obtain a multi-dimensional standard task coefficient set. Based on the multidimensional standard task coefficient set and the standard business requirement set, the multidimensional computing power requirement of the real-time business requirement is determined; Based on the business process information and the business type, obtain the corresponding business constraint information, wherein the business constraint information includes at least the upper bound of delay, the lower bound of reliability, and the security level; The multidimensional computing power demand and the business constraint information are vectorized and concatenated to form the real-time business components.

[0146] The pre-construction of the standard business requirements set includes: A predefined set of standard computing tasks is provided, which includes multiple benchmark computing tasks associated with parameterized computing features, and each benchmark computing task represents a type of computing mode. For each of the benchmark computing tasks, offline benchmark performance tests are performed on multiple computing nodes in the target computing power network to obtain the original performance characterization values ​​of the multiple computing nodes for each of the benchmark computing tasks. Using the statistical extreme values ​​of all the computing power nodes as a normalization benchmark, the original performance characterization values ​​are normalized to obtain the normalized performance characterization values ​​of each computing power node for each benchmark computing task. Establish a mapping relationship between business types and the standard computing task set. For each business type, determine the proportion coefficient of each benchmark computing task according to its typical execution process to form the standard business requirement set.

[0147] In one embodiment, the optimal route solving module 13 is used for: Based on the prior multidimensional evaluation rules of the multidimensional scheduling strategy and the target scenario, the weight coefficient of the multidimensional scheduling strategy is determined. The multidimensional scheduling strategy includes cost-aware scheduling strategy, load-aware scheduling strategy, energy-efficiency-aware scheduling strategy and service level agreement-aware scheduling strategy. Combining the multi-dimensional scheduling strategy, the weight coefficients of the multi-dimensional scheduling strategy, and the real-time service components, a multi-objective optimization task is constructed, wherein the objective function of the multi-objective optimization task is the weighted sum of the weight coefficients of the multi-dimensional scheduling strategy and the sub-objective function values ​​corresponding to each scheduling strategy; The search lower limit of the heterogeneous computing resource characteristic map is limited to the edge side, multiple alternative routing schemes are generated, and the task cost value of the multiple alternative routing schemes for the multi-objective optimization solution task is calculated and obtained. The optimal computation route is determined based on the task cost.

[0148] In one embodiment, the task allocation module 14 is used for: Update the heterogeneous computing resource characteristic map, and extract the feature embedding vectors of the path nodes using the optimal computing route as an index; Based on multiple feature embedding vectors, and using multiple computing power dimensions as coordinates, the multiple feature embedding vectors are mapped to the corresponding multi-dimensional space to construct the subnet multi-dimensional computing power spectrum and form the subnet multi-dimensional computing power spectrum space. Using the business constraint information in the real-time business components as hard constraints, unqualified end-side computing power nodes in the subnet multi-dimensional computing power spectrum space are screened out; The multidimensional computing power demand in the real-time service components is decomposed into the dimensional demand radius of each dimension in the multidimensional computing power spectrum space of the subnet. Determine whether there are edge computing power nodes above the required radius of multiple dimensions. If so, select the edge computing power node as the target task node using a greedy algorithm and output the edge-end task allocation result accordingly.

[0149] One of the cloud-edge-device collaborative scheduling methods for computing power networks also includes: If there are no edge computing nodes above the required radii of multiple dimensions, then a set of edge computing nodes is selected in the multidimensional computing power spectrum space of the subnet based on a greedy algorithm. Determine whether the sum of the edge computing power node set satisfies the required radius of multiple dimensions. If it does, output the edge computing power node set as the target task node set and output the edge-end task allocation result accordingly. Otherwise, tasks are split based on the real-time business components, and edge-end task allocation is performed on the multiple split tasks in the task splitting results.

[0150] In one embodiment, the business demand scheduling module 15 is used for: Multiple edge computing nodes perform local computations based on the edge-to-edge task allocation results and periodically report their execution status to the corresponding edge computing nodes. The execution status includes at least the task completion progress and resource utilization rate. The edge computing nodes update the subnet's multidimensional computing power spectrum based on the reported status. When the state change rate of an edge computing node exceeds a preset rematch threshold, real-time rematching based on spectral position sliding is triggered, including: If the load increases, it will first expand to the neighborhood of the current spectral position. If the neighborhood is insufficient, it will expand to higher-level spectral positions along the spectral axis. If the load decreases, higher-level spectral positions are released first, while lower-level spectral positions contract along the spectral axis. The real-time rematching results are merged with the status reporting results and aggregated to the cloud computing nodes to update the heterogeneous computing resource characteristic map.

[0151] Compared with existing technologies, this application first performs hybrid sensing acquisition of the target computing network through a cloud-edge-device generalized sensing layer, constructs a graph structure based on topological relationships, and then embeds and updates the graph neural network by combining node state parameters. This achieves a negative correlation between the acquisition cycle and the rate of change of historical states under the edge side, providing a unified and computable feature input basis for subsequent multi-objective optimization routing and task allocation.

[0152] Secondly, by pre-constructing a standard set of business requirements, a unified metric is established for heterogeneous computing power nodes. Then, the business type of real-time business requirements is identified, business process information is invoked, and multi-dimensional computing power requirements are determined. At the same time, business constraint information is extracted and finally vectorized and concatenated into real-time business components, reducing the computational latency of real-time scheduling, avoiding resource mismatch caused by matching a single indicator, and improving maintainability.

[0153] Furthermore, by determining the weight coefficients based on the multi-dimensional scheduling strategy and the prior multi-dimensional evaluation rules calibrated by the target scenario, a multi-objective optimization task is constructed to take into account multiple conflicting objectives and avoid the one-sidedness of single-objective optimization. The search lower limit is limited to generating alternative routing schemes on the edge side and calculating the task cost value. Finally, the optimal computation route is determined based on the cost value to avoid scheduling the task to overloaded or faulty nodes. At the same time, the comparison of task cost values ​​provides an explanatory basis for route selection.

[0154] Furthermore, by updating the heterogeneous computing resource characteristic map and extracting feature embedding vectors using the optimal computing route as an index, a multi-dimensional computing power spectrum of the subnet is constructed to form a spectral space, avoiding resource mismatch due to single-dimensional matching. Unqualified end-side nodes are screened out using business constraint information as hard constraints, and it is determined whether there is a single node that meets the requirements, ensuring that the end-side nodes meet the key requirements and guaranteeing the quality of business services. The multi-dimensional computing power requirement is decomposed into dimensional requirement radii, and an end-side node set is constructed. The end-side node set is selected through a greedy algorithm and summed for discrimination, realizing the multi-node aggregation of end-side computing power, breaking through the resource boundaries of a single subnet, while maintaining the correctness of data dependencies.

[0155] Ultimately, multiple edge nodes perform calculations based on the allocation results and report the execution status to the edge side, providing a data foundation for dynamic adjustments and achieving real-time observability of task execution. The edge side updates the subnet's multi-dimensional computing power spectrum based on the reported results, which can avoid selecting overloaded nodes in subsequent new task allocations. When the state change rate of an edge node exceeds a preset rematch threshold, real-time rematching based on spectrum position sliding is triggered. Finally, the rematching results and the status reporting results are merged and summarized in the cloud to update the heterogeneous computing resource characteristic map, avoiding decision-making biases caused by outdated information.

Claims

1. A cloud-edge-end collaborative scheduling method for a computing power network, characterized in that, include: Based on the topological relationship of the target computing power network, an initial heterogeneous computing resource characteristic map based on graph neural network is established, and the node state parameters obtained by the cloud-edge-device general perception layer are embedded and updated. Obtain real-time business requirements and, in conjunction with a pre-built set of standard business requirements, perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components. Obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, and combine the real-time service components with the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization and obtain the optimal computing route; Based on the optimal computing route and the heterogeneous computing resource characteristic map, a multi-dimensional computing power spectrum of the subnet is constructed, and edge-end task allocation is performed in conjunction with the real-time service components. By integrating the optimal computational route with the edge-to-end task allocation results, collaborative scheduling of real-time business requirements is performed.

2. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, Based on the topological relationships of the target computing power network, a heterogeneous computing resource characteristic map based on graph neural networks is initialized and established. This map is then embedded and updated using node state parameters obtained from the cloud-edge-device generalized perception layer, including: The cloud-edge-device generalized perception layer is used to perform hybrid perception and acquisition on the target computing network to obtain node status parameters. Based on the aforementioned topological relationship, a graph structure is constructed using the relative performance values ​​of computing power nodes and computing power nodes facing multiple preset standard computing tasks as graph nodes, and the network distance between computing power nodes as relation edges. Based on the node state parameters, the graph nodes and the relation edges are embedded and updated in the graph structure to generate feature embedding vectors of multiple graph nodes, thereby obtaining the heterogeneous computing resource characteristic map. The feature embedding vector is defined as the embedding vector of itself updated by a nonlinear transformation function after weighting the neighbor information of neighboring nodes with edge weights.

3. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 2, wherein, Through the cloud-edge-device generalized sensing layer, hybrid sensing data is collected from the target computing network to obtain node state parameters, including: The cloud-edge-device pervasive perception layer includes a cloud center perception agent, an edge-side perception agent, and a device-side perception agent. The hybrid perception acquisition includes event-driven acquisition and adaptive periodic acquisition. The cloud center perception agent and the edge side perception agent perform adaptive periodic data collection, wherein the collection period of the adaptive periodic data collection is negatively correlated with the historical state change rate under the edge side computing power node. The edge-side sensing agent and the endpoint-side sensing agent synchronously perform the adaptive periodic acquisition and the event-driven acquisition.

4. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, Obtain real-time business requirements and, in conjunction with a pre-built set of standard business requirements, perform standardized requirement decomposition on these requirements to obtain real-time business components, including: Identify the business type of the real-time business requirement, and invoke business process information based on the business type; By combining the business process information with the standard business requirement set, the real-time business requirements are decomposed into standardized requirements to obtain a multi-dimensional standard task coefficient set. Based on the multidimensional standard task coefficient set and the standard business requirement set, the multidimensional computing power requirement of the real-time business requirement is determined; Based on the business process information and the business type, obtain the corresponding business constraint information, wherein the business constraint information includes at least the upper bound of delay, the lower bound of reliability, and the security level; The multidimensional computing power demand and the business constraint information are vectorized and concatenated to form the real-time business components.

5. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, The pre-built standard business requirements set includes: A predefined set of standard computing tasks is provided, which includes multiple benchmark computing tasks associated with parameterized computing features, and each benchmark computing task represents a type of computing mode. For each of the benchmark computing tasks, offline benchmark performance tests are performed on multiple computing nodes in the target computing power network to obtain the original performance characterization values ​​of the multiple computing nodes for each of the benchmark computing tasks. Using the statistical extreme values ​​of all the computing power nodes as a normalization benchmark, the original performance characterization values ​​are normalized to obtain the normalized performance characterization values ​​of each computing power node for each benchmark computing task. Establish a mapping relationship between business types and the standard computing task set. For each business type, determine the proportion coefficient of each benchmark computing task according to its typical execution process to form the standard business requirement set.

6. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, Obtain the multi-dimensional scheduling strategy weight coefficients for the target scenario, and combine them with the real-time service components and the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization, thereby obtaining the optimal computing route, including: Based on the prior multidimensional evaluation rules of the multidimensional scheduling strategy and the target scenario, the weight coefficient of the multidimensional scheduling strategy is determined. The multidimensional scheduling strategy includes cost-aware scheduling strategy, load-aware scheduling strategy, energy-efficiency-aware scheduling strategy and service level agreement-aware scheduling strategy. Combining the multi-dimensional scheduling strategy, the weight coefficients of the multi-dimensional scheduling strategy, and the real-time service components, a multi-objective optimization task is constructed, wherein the objective function of the multi-objective optimization task is the weighted sum of the weight coefficients of the multi-dimensional scheduling strategy and the sub-objective function values ​​corresponding to each scheduling strategy; The search lower limit of the heterogeneous computing resource characteristic map is limited to the edge side, multiple alternative routing schemes are generated, and the task cost value of the multiple alternative routing schemes for the multi-objective optimization solution task is calculated and obtained. The optimal computation route is determined based on the task cost.

7. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, Based on the optimal computational route and the heterogeneous computing resource characteristic map, a multi-dimensional computing power spectrum for the subnet is constructed, and edge-to-end task allocation is performed in conjunction with the real-time service components, including: Update the heterogeneous computing resource characteristic map, and extract the feature embedding vectors of the path nodes using the optimal computing route as an index; Based on multiple feature embedding vectors, and using multiple computing power dimensions as coordinates, the multiple feature embedding vectors are mapped to the corresponding multi-dimensional space to construct the subnet multi-dimensional computing power spectrum and form the subnet multi-dimensional computing power spectrum space. Using the business constraint information in the real-time business components as hard constraints, unqualified end-side computing power nodes in the subnet multi-dimensional computing power spectrum space are screened out; The multidimensional computing power demand in the real-time service components is decomposed into the dimensional demand radius of each dimension in the multidimensional computing power spectrum space of the subnet. Determine whether there are edge computing power nodes above the required radius of multiple dimensions. If so, select the edge computing power node as the target task node using a greedy algorithm and output the edge-end task allocation result accordingly.

8. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 7, wherein, Also includes: If there are no edge computing nodes above the required radii of multiple dimensions, then a set of edge computing nodes is selected in the multidimensional computing power spectrum space of the subnet based on a greedy algorithm. Determine whether the sum of the edge computing power node set satisfies the required radius of multiple dimensions. If it does, output the edge computing power node set as the target task node set and output the edge-end task allocation result accordingly. Otherwise, tasks are split based on the real-time business components, and edge-end task allocation is performed on the multiple split tasks in the task splitting results.

9. The cloud-edge-end collaborative scheduling method for a computing power oriented network of claim 1, wherein, After integrating the optimal computational route with the edge-to-end task allocation results, and performing collaborative scheduling of real-time service requirements, the process further includes: Multiple edge computing nodes perform local computations based on the edge-to-edge task allocation results and periodically report their execution status to the corresponding edge computing nodes. The execution status includes at least the task completion progress and resource utilization rate. The edge computing nodes update the subnet's multidimensional computing power spectrum based on the reported status. When the state change rate of an edge computing node exceeds a preset rematch threshold, real-time rematching based on spectral position sliding is triggered, including: If the load increases, it will first expand to the neighborhood of the current spectral position. If the neighborhood is insufficient, it will expand to higher-level spectral positions along the spectral axis. If the load decreases, higher-level spectral positions are released first, while lower-level spectral positions contract along the spectral axis. The real-time rematching results are merged with the status reporting results and aggregated to the cloud computing nodes to update the heterogeneous computing resource characteristic map.

10. A cloud-edge-end collaborative scheduling system for a computing power network, characterized in that, The system is used to implement the cloud-edge-device collaborative scheduling method for computing power networks according to any one of claims 1-9, the system comprising: The embedding update module is used to initialize and establish a heterogeneous computing resource characteristic map based on the topological relationship of the target computing power network, and to perform embedding update in combination with the node state parameters obtained by the cloud-edge-device general perception layer. The business requirement decomposition module is used to obtain real-time business requirements and, in conjunction with a pre-built set of standard business requirements, to perform standardized requirement decomposition on the real-time business requirements to obtain real-time business components. The optimal route solving module is used to obtain the multi-dimensional scheduling strategy weight coefficients of the target scenario, and combine the real-time service components with the heterogeneous computing resource characteristic map to perform optimal route solving based on multi-objective optimization to obtain the optimal computing route. The task allocation module is used to construct a multi-dimensional computing power spectrum of the subnet based on the optimal computing route and the heterogeneous computing resource characteristic map, and to perform edge-end task allocation in combination with the real-time service components. The business demand scheduling module is used to integrate the optimal computing route with the edge-end task allocation results to perform real-time collaborative scheduling of business demands.

Images