Intelligent fusion terminal cloud edge resource elastic scheduling method and system for low latency scenario

Abstract

This invention provides a method and system for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios, relating to the field of resource scheduling technology. The method includes: acquiring and decomposing the end-to-end latency constraints of a target low-latency service request, and setting N initial candidate execution paths; mapping the N initial candidate execution paths to path state vectors and performing multi-dimensional evaluation analysis; performing path iterative search of the N initial candidate execution paths based on the comprehensive evaluation vector; configuring an elastic scheduling strategy using the path iterative search results; and performing intelligent scheduling management of the target low-latency service request. This invention solves the technical problems of existing cloud-edge-device scheduling methods that rely on static rules or single-node load, leading to unstable latency and low resource scheduling efficiency for low-latency services. It achieves the technical effect of improving the latency stability and resource scheduling efficiency of low-latency service requests through multi-dimensional constraint decomposition and path iterative optimization to realize cloud-edge-device collaborative scheduling.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, specifically to a method and system for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios. Background Technology

[0002] With the coordinated development of cloud computing, edge computing, and terminal devices, low-latency business scenarios such as industrial control, real-time video processing, and AR / VR interaction have imposed strict constraints on end-to-end latency, requiring rapid resource coordination and task scheduling between the cloud, edge, and terminal. Existing technologies often employ scheduling methods based on static rules or dynamic scheduling methods based on the load status of a single node. For example, task offloading decisions are made based on edge node CPU utilization or cloud queue length, or fixed task allocation strategies are pre-set to achieve resource scheduling control. Some solutions also introduce simple load balancing mechanisms to distribute resources evenly among nodes.

[0003] However, the aforementioned methods typically only consider a single dimension of resource status, lacking unified modeling and analysis capabilities for end-to-end latency constraints, network transmission overhead, and multi-node coupling relationships. This makes it difficult to reflect the overall impact during cloud-edge-device collaborative execution. Furthermore, the lack of a collaborative optimization mechanism between the service execution process and the operational characteristics of the communication network can easily lead to insufficient adaptability of scheduling strategies under network fluctuations or load changes, resulting in large latency fluctuations for low-latency services, uneven resource utilization, and low overall scheduling efficiency.

[0004] In summary, existing technologies suffer from technical problems such as unstable latency for low-latency services and low resource scheduling efficiency due to reliance on static rules or single-node load in cloud-edge scheduling methods. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios, in order to solve the technical problems of existing cloud-edge scheduling methods that rely on static rules or single node load, resulting in unstable latency of low-latency services and low resource scheduling efficiency.

[0006] In view of the above problems, this application provides a method and system for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios.

[0007] The first aspect of this application provides a method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios. This method includes: obtaining end-to-end latency constraints of a target low-latency service request; decomposing the end-to-end latency constraints into terminal processing constraints, edge transmission constraints, and computational constraints; constructing a multi-dimensional evaluation constraint space corresponding to the decomposition results; setting N initial candidate execution paths using the multi-dimensional evaluation constraint space; mapping the N initial candidate execution paths to path state vectors and inputting them into a multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis, outputting a comprehensive evaluation vector; performing path iterative search on the N initial candidate execution paths based on the comprehensive evaluation vector, establishing path iterative search results; configuring an elastic scheduling strategy using the path iterative search results; and performing intelligent scheduling management of the target low-latency service request based on the elastic scheduling strategy.

[0008] Optionally, the multidimensional evaluation constraint space is used to characterize the constraint boundaries of different execution paths in terms of latency, load consumption, and network transmission overhead.

[0009] Optionally, based on the coupling relationship between constraints in the multidimensional evaluation constraint space, a convergence potential energy distribution model of the feasible path region is constructed. The convergence potential energy distribution model is used to evaluate the trend strength of different cloud, edge, and terminal combination execution paths in satisfying the convergence of the target low-latency service request. The potential energy gradient of the cloud-side execution path, edge-side execution path, and terminal-side execution path is calculated in the convergence potential energy distribution model to determine the convergence attraction strength of different path combinations in the feasible region. Based on the convergence attraction strength, N initial candidate execution paths with convergence guidance characteristics are generated in the feasible path region.

[0010] Optionally, cloud-side resource nodes, edge resource nodes, and terminal resource nodes are divided into an ordered hierarchical structure according to task processing stages. Based on the computing power, current load status, and network connectivity of each node, an initial resource association graph containing node attributes and inter-node transmission overhead is constructed. The convergent potential energy distribution model is mapped to the initial resource association graph, assigning potential energy values ​​to the nodes in the graph. Based on the latency contribution, bandwidth usage, and computation queuing latency between nodes, the potential energy change of the edge weights between nodes is calculated, forming a weighted path generation graph with potential energy gradient constraints. In the weighted path generation graph, terminal resource nodes are used as... Starting from a source, cloud-side resource node, or edge resource node, a constrained path expansion is performed along the direction of decreasing potential energy to generate a set of candidate paths that meet convergence constraints. The number of effective surviving branches of the path expansion is counted, and an adaptive adjustment function for path size is constructed based on the decay rate of the number of effective paths with expansion depth. The number of paths corresponding to the convergence and stability interval of the adaptive adjustment function is used as the number of candidate execution paths N. The generated set of candidate paths is subjected to multi-dimensional redundancy removal screening based on total path delay, resource occupancy balance, and network overhead, and path combinations that meet the difference threshold are retained as N initial candidate execution paths.

[0011] Optionally, the multi-dimensional evaluation sub-channel includes a delay sub-channel, a load sub-channel, and a transmission sub-channel. The delay sub-channel outputs a delay sub-score based on the queuing delay and processing delay at each stage of the path. The load sub-channel constructs a resource consumption mapping relationship based on node computing power and real-time load status and outputs a resource consumption sub-score. The transmission sub-channel outputs a network transmission cost sub-score based on link bandwidth, congestion level, and transmission stability. After each sub-channel completes the calculation of all sub-scores for the current round, the corresponding state sensitivity factor is extracted. The state sensitivity factor is used to characterize the gradient of the sub-score's impact on the overall path performance. The sub-scores and state sensitivity factors are transmitted to the other sub-channels through an inter-channel broadcast mechanism. Based on the state sensitivity factor, a weighted correction process is performed on the corresponding feature components in the path state vector, so that the path state vector characterizes the coupling effect of cross-dimensional constraints in the next round of input, forming an updated path state vector after multi-sub-channel interactive correction. Based on the updated path state vector, multiple rounds of sub-channel evaluation and inter-channel interactive correction are repeatedly performed until the changes in all sub-scores output by each sub-channel meet the preset convergence condition. After fusion processing, a comprehensive evaluation vector is output to characterize the comprehensive performance of the candidate path.

[0012] Optionally, a search reference direction is established based on the comprehensive evaluation vector and N initial candidate execution paths; after performing a search perturbation on the search reference direction, search optimization is performed under the constraint of the execution direction; and path iterative search is performed on the execution path based on the search optimization result.

[0013] Optionally, based on the search reference direction, each candidate execution path is mapped to the corresponding search direction space, and a search priority is configured for each candidate execution path according to the comprehensive evaluation vector; after search perturbation, multiple neighborhood perturbation paths are configured for each candidate execution path, and the neighborhood perturbation paths are constructed by locally replacing or adjusting the node selection, task allocation ratio, or transmission link in the path; the updated path state vector of each neighborhood perturbation path is calculated, the updated path state vector is used to perform effectiveness evaluation, and the effectiveness evaluation results are used for search optimization.

[0014] Optionally, the elastic scheduling strategy can be distributed to the corresponding cloud-side resource nodes and edge resource nodes to perform resource elastic scheduling management.

[0015] Optionally, during the generation process, the elastic scheduling strategy performs redundant backup configuration of the resource allocation scheme corresponding to the candidate execution path. When the performance degradation of the main path meets the predetermined conditions during execution, it switches to the alternative path. The switching of the alternative path is triggered based on the real-time monitored latency deviation.

[0016] A second aspect of this application provides an intelligent converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios. The system comprises: a constraint space construction module, used to obtain end-to-end latency constraints of a target low-latency service request, decompose the end-to-end latency constraints into terminal processing constraints, edge transmission constraints, and computational constraints, and construct a multi-dimensional evaluation constraint space corresponding to the decomposition results; a candidate path setting module, used to set N initial candidate execution paths using the multi-dimensional evaluation constraint space; an evaluation analysis module, used to map the N initial candidate execution paths into path state vectors, input them into a multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis, and output a comprehensive evaluation vector; a path search module, used to perform iterative path search of the N initial candidate execution paths based on the comprehensive evaluation vector, and establish iterative path search results; and a scheduling management module, used to configure an elastic scheduling strategy using the iterative path search results, and perform intelligent scheduling management of the target low-latency service request according to the elastic scheduling strategy.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0018] The method provided in this application obtains the end-to-end latency constraints of the target low-latency service request, decomposes the end-to-end latency constraints into terminal processing constraints, edge transmission constraints, and computational constraints, and constructs a multi-dimensional evaluation constraint space corresponding to the decomposition results; uses the multi-dimensional evaluation constraint space to set N initial candidate execution paths; maps the N initial candidate execution paths to path state vectors, inputs them into a multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis, and outputs a comprehensive evaluation vector; performs path iterative search on the N initial candidate execution paths according to the comprehensive evaluation vector to establish path iterative search results; configures an elastic scheduling strategy using the path iterative search results, and performs intelligent scheduling management of the target low-latency service request according to the elastic scheduling strategy. This achieves the technical effect of improving the latency stability and resource scheduling efficiency of low-latency service requests by realizing cloud-edge-device collaborative scheduling through multi-dimensional constraint decomposition and path iterative optimization.

[0019] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios provided in this application.

[0022] Figure 2 A schematic diagram of the structure of the intelligent converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios provided in this application.

[0023] Figure labeling: Constraint space construction module 11, Candidate path setting module 12, Evaluation and analysis module 13, Path search module 14, Scheduling management module 15. Detailed Implementation

[0024] This application provides a method and system for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios. It addresses the technical problems of existing cloud-edge scheduling methods that rely on static rules or single-node load, leading to unstable latency and low resource scheduling efficiency for low-latency services. The method achieves improved latency stability and resource scheduling efficiency for low-latency service requests through multi-dimensional constraint decomposition and iterative path optimization to realize collaborative scheduling between cloud, edge, and terminal.

[0025] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0026] Example 1, as Figure 1 As shown, this application provides a method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios. The method includes: Obtain the end-to-end latency constraints of the target low-latency service request, decompose the end-to-end latency constraints into terminal processing constraints, edge transmission constraints, and computation constraints, and construct a multi-dimensional evaluation constraint space corresponding to the decomposition results.

[0027] Furthermore, the multidimensional evaluation constraint space is used to characterize the constraint boundaries of different execution paths in terms of latency, load consumption, and network transmission overhead.

[0028] Specifically, the system obtains the target low-latency service request from the request access layer of the service scheduling system, and parses the end-to-end latency constraint from the service quality description field or SLA protocol of the target low-latency service request. The end-to-end latency constraint is explicitly carried by the service initiator in the request message, or it is mapped by the service orchestration module in the policy library according to the service type, such as real-time video, industrial control instructions, AR interaction, etc. The end-to-end latency constraint represents the maximum allowable time limit for a complete service request from the terminal to the cloud-edge-device collaborative processing and return of the result, and its unit is milliseconds.

[0029] Based on end-to-end latency constraints, and combined with business data characteristics and real-time operational status, segmented analysis is performed. First, business requests are structured and parsed to extract latency-related basic parameters, including: input data volume such as packet size and frame size; task type such as inference computation and encoding processing; and interaction mode such as single request or continuous streaming request. These latency-related basic parameters are obtained through application layer protocol fields, task description headers, or terminal SDK reporting. Simultaneously, the current terminal processing capacity, network link status, and average computation time of edge / cloud nodes are monitored. Current terminal processing capacity includes unit data processing time; network link status includes bandwidth and round-trip latency; and the average computation time of edge / cloud nodes is obtained through historical task statistics, yielding average queuing time and execution time.

[0030] Based on this, the end-to-end latency constraints are decomposed into achievable capabilities for each stage: The base processing time on the terminal side is calculated based on the unit data processing time (e.g., 2ms per MB of data) and the current task data volume. This base processing time is then adjusted according to the terminal's current load, such as CPU utilization; for example, the base processing time is amplified when the load is high, thus obtaining the terminal processing constraint. The pure transmission time is calculated based on the data volume and current link bandwidth. This is then superimposed with the real-time measured network round-trip latency and the average computation time of edge / cloud nodes to obtain the base time for the network stage. The edge transmission constraint is determined accordingly. The remaining latency budget is allocated to the computation constraint. The rationality of the computation constraint is verified by combining the current average queuing time and unit computation time of edge and cloud nodes. If the computation constraint is insufficient to support task execution, the allocation ratio of the first two parts is adjusted in reverse, for example, by compressing the edge transmission constraint or the terminal processing constraint, until the basic executable conditions are met.

[0031] After obtaining the terminal processing constraints, edge transmission constraints, and computation constraints, these are used as the latency constraint boundaries. Load consumption constraints and network transmission overhead constraints are also introduced. The load consumption constraint is determined by summarizing the CPU utilization increments and queue length changes of nodes involved in the path, such as terminals, edges, or clouds. For example, for a single node, the CPU utilization increments and queue length changes are normalized and then weighted to obtain the node load consumption value. The weights are preset according to the service type; for example, real-time services increase the weight of CPU utilization increments, while batch processing increases the weight of queue length changes. The load consumption values ​​of all nodes in the path are accumulated to form a path-level load consumption index, which is then normalized and mapped as a load consumption constraint value to reflect the overall pressure level of the path on the overall computing resources. Normalization can be achieved through min-max normalization or z-score standardization.

[0032] Network transmission overhead constraints are calculated by comprehensively considering the bandwidth ratio occupied by the path, link congestion level, and retransmission rate. First, for each link involved in the path, such as terminal-edge or edge cloud, three core indicators are collected: link bandwidth occupancy ratio (current service bandwidth / available link bandwidth), link congestion level classification based on queue occupancy rate or ECN tagging, and data transmission retransmission rate, used to characterize link reliability and transmission loss. For a single link, the link bandwidth occupancy ratio, link congestion level classification, and data transmission retransmission rate are normalized. The congestion level can be mapped from discrete levels to continuous values. Similarly, the link network overhead value is calculated by weighted summation, accumulating the network overhead of all links in the path to obtain the path-level network transmission overhead, which is then normalized to the 0-1 range, thus serving as the network transmission overhead constraint to characterize the overall communication cost and stability burden of the path during data transmission.

[0033] Furthermore, a multi-dimensional evaluation constraint space is constructed based on latency constraints, load consumption constraints, and network transmission overhead constraints. This multi-dimensional evaluation constraint space is used to characterize the constraint boundaries of different execution paths in terms of latency, load consumption, and network transmission overhead, thereby enabling a more accurate grasp of the latency requirements of different stages of the business and better meeting the needs of low-latency services during resource scheduling.

[0034] N initial candidate execution paths are set using the multidimensional evaluation constraint space.

[0035] Furthermore, N initial candidate execution paths are set using the multidimensional evaluation constraint space, including: constructing a convergence potential energy distribution model of the feasible region of the path based on the coupling relationship between the constraints in the multidimensional evaluation constraint space, wherein the convergence potential energy distribution model is used to evaluate the trend strength of different cloud, edge, and terminal combination execution paths in meeting the convergence of the target low-latency service request; calculating the potential energy gradient of the cloud-side execution path, edge-side execution path, and terminal-side execution path in the convergence potential energy distribution model to determine the convergence attraction strength of different path combinations in the feasible region; and generating N initial candidate execution paths with convergence guidance characteristics within the feasible region based on the convergence attraction strength.

[0036] Specifically, based on the multidimensional evaluation constraint space, each candidate path is represented as a three-dimensional constraint vector P=(D,L,N), where D is the latency constraint, L is the load consumption constraint, and N is the network overhead constraint. D, L, and N are all normalized constraint occupancy ratios, with values ​​ranging from [0,1]. For example, D = actual latency / SLA latency limit, L = actual load increment / node capacity limit, and N = actual network overhead / link capacity limit. Based on historical scheduling data and real-time operational data, the coupling relationships between constraints in the multidimensional evaluation constraint space are statistically analyzed. For instance, excessive terminal processing time may increase overall latency, thus affecting the time allocation for edge transmission and computation. Simultaneously, excessive terminal load will also affect its processing capacity, further impacting latency. Excessive bandwidth occupancy in network transmission overhead may lead to increased data transmission latency, affecting the time arrangement for edge transmission and computation.

[0037] Based on the coupling relationships between various constraints, a convergence potential energy distribution model of the feasible path region is constructed. First, the ideal convergence target point P is defined. =(1,0,0), where D=1 indicates that the latency upper limit is just met, and L and N approaching 0 indicate that resource and network overhead are minimized. Then, a path deviation function is constructed, using the distance between the path vector in the multidimensional constraint space and the target point as the source of basic potential energy, for example, using weighted Euclidean distance: α, β, and γ are used to reflect the importance weights of latency, load, and network. α is the highest in low-latency scenarios. For example, let α be set to 0.5, and β and γ be 0.25 respectively. E(P) represents the potential energy value of a path that is closer to the ideal state of low latency, low load, and low network overhead.

[0038] Based on this, a coupling correction term is introduced to reflect the mutual influence between constraints. For example, based on the coupling relationship between load consumption constraints and delay constraints and the coupling relationship between network overhead constraints and delay constraints, an extended potential energy function is constructed: E′(P)=E(P)+δ1·D L+δ2·D·N, where D·L represents the coupling cost of increasing computational load to reduce latency, D·N represents the coupling cost of increasing transmission cost to reduce latency, and δ1 and δ2 are coupling strength coefficients used to reflect the degree of cloud-edge collaboration in the system. δ1=1-edge computing capability / system maximum computing capability, δ2=1-average bandwidth utilization / link peak bandwidth. The stronger the edge capability, the smaller δ1. By extending the potential energy function, the potential energy value corresponding to each point in the multidimensional evaluation constraint space is calculated, thereby constructing a convergence potential energy distribution model of the feasible path region. This model is used to evaluate the trend strength of different cloud, edge, and terminal combinations in meeting the convergence of low-latency service requests. The smaller the potential energy value E′(P), the closer the path is to the convergence direction that meets the low-latency target.

[0039] After the potential energy distribution model is constructed, potential energy gradients are calculated for cloud-side execution paths, edge-side execution paths, and terminal-side execution paths. The potential energy gradient refers to the direction and magnitude of the convergence potential energy change of a node or path combination within the feasible region of the path when a unit structural change occurs, such as adding / removing an edge node or adjusting the task offloading ratio. This is obtained by differencing the state vectors of adjacent paths, for example, by calculating ΔE / ΔP, where ΔE is the potential energy change and ΔP is the path structural change. This yields the convergence attraction strength of each candidate path within the feasible region. A greater convergence attraction strength indicates that the path is more likely to quickly meet low-latency constraints and has a more favorable resource cost.

[0040] After obtaining the convergence attraction strength of each path, N initial candidate execution paths with convergence guidance characteristics are generated within the feasible region of the path using methods such as random sampling or heuristic search. For example, starting from the terminal node and with cloud nodes or edge nodes as potential endpoints, the search is expanded within the feasible region according to the principles of prioritizing high potential energy and low cost. In each step of path expansion, the node or link with the largest decrease in potential energy is selected first, thus forming a path generation process with convergence guidance characteristics. Based on the convergence attraction strength, N initial candidate execution paths are generated. Structurally, all N initial candidate execution paths are located on paths with a significant decreasing potential energy trend, i.e., they have convergence guidance characteristics and are more likely to be close to the optimal solution that satisfies the target low-latency service request. N is a positive integer representing the number of initial candidate execution paths.

[0041] By constructing a convergent potential energy distribution model, the originally disordered cloud-edge-device combined path search is transformed into a potential energy descent process with directional guidance. This introduces optimization preferences at the candidate path generation stage, which can more effectively narrow the search range, improve search efficiency, and thus find the execution path that meets business requirements more quickly, thereby improving scheduling efficiency.

[0042] Furthermore, generating N initial candidate execution paths with convergence guidance characteristics within the feasible path domain also includes: dividing cloud-side resource nodes, edge resource nodes, and terminal resource nodes into an ordered hierarchical structure according to task processing stages, and constructing an initial resource association graph containing node attributes and inter-node transmission overhead based on the computing power, current load status, and network connectivity of each node; mapping the convergence potential energy distribution model onto the initial resource association graph, assigning potential energy values ​​to the nodes in the graph, and calculating the potential energy change of the edge weights between nodes based on the latency contribution, bandwidth occupancy, and computation queuing latency between nodes, forming a weighted path generation graph with potential energy gradient constraints; in the... In the weighted path generation graph, starting from the terminal resource node and ending at the cloud-side resource node or edge resource node, constrained path expansion is performed along the direction of decreasing potential energy to generate a set of candidate paths that meet convergence constraints. The number of effective surviving branches of the path expansion is counted, and an adaptive adjustment function for path size is constructed based on the decay rate of the number of effective paths with the expansion depth. The number of paths corresponding to the convergence and stability interval of the adaptive adjustment function is used as the number N of candidate execution paths. The generated set of candidate paths is subjected to multi-dimensional redundancy removal screening based on total path delay, resource occupancy balance, and network overhead, and path combinations that meet the difference threshold are retained as N initial candidate execution paths.

[0043] Specifically, determining the initial candidate execution path can also involve dividing cloud-side resource nodes, edge resource nodes, and terminal resource nodes into an ordered hierarchical structure according to the task processing stage. The ordered hierarchical structure refers to dividing the nodes into the terminal layer, edge layer, and cloud layer according to the business execution order. For example, terminal resource nodes are responsible for data collection and preliminary processing and are classified as the terminal layer; edge resource nodes further process and analyze the data uploaded by the terminal and are classified as the edge layer; and cloud-side resource nodes are responsible for large-scale data storage, complex calculations, and global decision-making and are classified as the cloud layer.

[0044] Real-time status data of cloud-side resource nodes, edge resource nodes, and terminal resource nodes are obtained through a resource scheduling system. This data includes computing power, current load status, and network connectivity. Computing power refers to CPU or GPU computing power metrics and unit task processing rate. Current load status refers to CPU utilization and task queue length. Network connectivity refers to link reachability, RTT latency, and bandwidth capacity. Based on the computing power, current load status, and network connectivity of each node, an initial resource association graph is constructed, containing node attributes and inter-node transmission overhead. A node represents a specific cloud, edge, or terminal resource instance; node attributes are represented as a vector containing information such as unit data transmission latency, bandwidth utilization, and link congestion probability; and edges represent reachable communication links between nodes.

[0045] The convergent potential energy distribution model is mapped to the initial resource association graph, and the nodes in the initial resource association graph are assigned corresponding potential energy values. For example, for node i in the initial resource association graph, its potential energy value Ei is determined according to its position in the convergent potential energy distribution model. Based on the latency contribution, bandwidth usage, and calculation queuing delay between nodes, the potential energy change of the edge weight between nodes is calculated. The edge weight between node i and node j is wij, and its potential energy change ΔEij is calculated using the formula ΔEij=w1×dij+w2×bij+w3×qij, where dij is the latency contribution from node i to node j, bij is the bandwidth usage, qij is the calculation queuing delay, and w1, w2, and w3 are weight coefficients that can be adjusted according to business requirements. Edge weight wij represents the basic transmission and scheduling cost between node i and node j. Historical link statistics between node i and node j are collected through a resource scheduling system, including average network latency, average bandwidth utilization, and average queuing time. These three metrics are normalized and then weighted and fused to determine the weights. The weights can be set based on historical data using the Analytic Hierarchy Process (AHP) or based on expert experience. Based on this, the edge weights between nodes are updated to wij′ = wij + ΔEij, forming a weighted path generation graph with potential gradient constraints.

[0046] Starting from terminal resource nodes and using edge nodes or cloud nodes as potential endpoints, a constrained path expansion search is performed in the weighted path generation graph. During the path search, a potential energy decreasing priority strategy is adopted, meaning that in each expansion step, neighboring nodes with smaller potential energy values ​​and negative edge weight potential energy changes are prioritized for expansion. Simultaneously, constraints such as load status and network connectivity are considered to avoid selecting nodes with excessive load or unreachable networks, thus preventing invalid path expansion. Through continuous expansion, a set of candidate paths satisfying convergence constraints is generated. Statistical analysis is performed on the number of effective surviving branches during the path expansion process, where an effective surviving branch is a path branch that can continue to expand to the endpoint under the constraints.

[0047] By analyzing the decay trend of the number of effective branches with increasing expansion depth, an adaptive adjustment function for path size is constructed. For example, an exponential decay function is used to describe the path space convergence characteristics, i.e., the number of effective paths N. eff (d) The variation of the extension depth d satisfies the exponential decay model. Where N0 is the initial number of effective paths, and λ is the decay coefficient, ranging from 0.1 to 1.5. By adjusting the value of λ, the number of paths corresponding to the path size adaptive adjustment function within the convergence and stability interval is taken as the number of candidate execution paths N. The convergence and stability interval can be determined by analyzing the changing trend of the path size adaptive adjustment function. For example, when the rate of change of the function value is less than the rate of change threshold, it is considered to have entered the convergence and stability interval. The rate of change threshold is determined by statistically analyzing the convergence rate of change sequence of all tasks from historical scheduling data and taking the 10th percentile of the historical rate of change.

[0048] The generated candidate path set undergoes multi-dimensional redundancy removal screening. Screening criteria include total path latency, resource utilization balance, and network transmission overhead. Only path combinations that meet these thresholds are retained. A difference threshold of 0.8 is set, and the Jaccard similarity coefficient is used to calculate the path similarity of the path set that meets the screening criteria. If the path similarity is greater than or equal to the difference threshold of 0.8, it is removed, retaining only the path set with significant structural differences and excellent performance. Finally, N initial candidate execution paths are output.

[0049] By graphing the cloud-edge-device resource structure and introducing a potential gradient guidance mechanism, along with adaptive path size control and multi-dimensional redundancy removal filtering, it is possible to generate initial candidate execution paths with convergence guidance characteristics and satisfying multiple constraints in complex cloud, edge, and device resource environments, thereby further improving the efficiency and quality of path search.

[0050] After mapping the N initial candidate execution paths to path state vectors, the vectors are input into the multidimensional evaluation sub-channel to perform multidimensional evaluation analysis and output a comprehensive evaluation vector.

[0051] Furthermore, the input is fed into a multidimensional evaluation sub-channel to perform multidimensional evaluation analysis, including: the multidimensional evaluation sub-channel includes a delay sub-channel, a load sub-channel, and a transmission sub-channel; the delay sub-channel is used to output a delay sub-score based on the queuing delay and processing delay at each stage of the path; the load sub-channel constructs a resource consumption mapping relationship based on node computing power and real-time load status and outputs a resource consumption sub-score; the transmission sub-channel outputs a network transmission cost sub-score based on link bandwidth, congestion level, and transmission stability; after each sub-channel completes the calculation of all sub-scores for the current round, the corresponding state sensitivity factor is extracted, and the state sensitivity factor is used to characterize... The gradient of the sub-score's impact on the overall path performance is transmitted to the remaining sub-channels via an inter-channel broadcast mechanism. Based on the state sensitivity factor, a weighted correction process is performed on the corresponding feature components in the path state vector, so that the path state vector represents the coupling effect of cross-dimensional constraints in the next round of input, forming an updated path state vector after multi-sub-channel interactive correction. Based on the updated path state vector, multiple rounds of sub-channel evaluation and inter-channel interactive correction are repeatedly performed until the changes in all sub-scores output by each sub-channel meet the preset convergence condition. After fusion processing, a comprehensive evaluation vector is output to characterize the overall performance of the candidate path.

[0052] Specifically, after obtaining N initial candidate execution paths, for any initial candidate execution path, its processing latency, node load information, and link status information at the terminal, edge, and cloud stages are extracted and normalized to construct a path state vector. The path state vector is a multi-dimensional vector that numerically represents the key features involved in the path execution process. Then, the path state vector is input into the multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis.

[0053] The multi-dimensional evaluation sub-channels include a delay sub-channel, a load sub-channel, and a transmission sub-channel. The delay sub-channel calculates and evaluates based on queuing delay and processing delay at each stage of the path. It extracts queuing delay and processing delay information related to each stage of the path from the path state vector, and outputs a delay sub-score based on the extracted queuing delay and processing delay using a specific delay evaluation algorithm, such as weighted average delay. The delay sub-score reflects the path's performance in terms of time delay. The load sub-channel constructs a resource consumption mapping relationship based on the computing power of each node in the path and its current load state. It obtains a resource consumption sub-score by weighting the CPU utilization increment with the queue length. The resource consumption sub-score reflects the path's performance in terms of node resource utilization. The transmission sub-channel extracts information such as link bandwidth occupancy, congestion level, and retransmission rate from the path state vector. Based on an appropriate network transmission cost evaluation method, such as a cost calculation method considering bandwidth utilization and congestion probability, it outputs a network transmission cost sub-score, which reflects the path's performance in terms of network transmission. After each round of sub-score calculation is completed, the corresponding state sensitivity factor is extracted. The state sensitivity factor refers to the gradient of the influence of a certain sub-channel score on the overall path performance change. It can be calculated by local perturbation, that is, applying a small change of a preset magnitude to a certain feature component in the path state vector, such as increasing unit load or delay, and recalculating the corresponding sub-channel score. Based on the ratio of the score difference before and after the perturbation to the perturbation magnitude, the change rate of the feature component with respect to the current sub-score is obtained, which serves as the state sensitivity factor to characterize the influence strength of this dimension on path performance.

[0054] Through an inter-channel broadcast mechanism, the sub-scores and corresponding state sensitivity factors of each sub-channel are transmitted to other sub-channels. This inter-channel broadcast mechanism is a data sharing method that ensures that each sub-channel can obtain relevant information from other sub-channels. Based on the received state sensitivity factors, a weighted correction process is performed on the corresponding feature components in the path state vector. That is, each sub-channel adjusts the weights of the feature components related to its own evaluation in the path state vector according to the state sensitivity factors received from other sub-channels. For example, if the load sub-channel receives a large state sensitivity factor from the delay sub-channel, it indicates that delay has a significant impact on the overall path performance. In this case, when correcting the path state vector, the load sub-channel will appropriately increase the weights of the feature components related to delay, so that the path state vector can represent the coupling effect of cross-dimensional constraints in the next round of input. That is, it considers the interrelationship between evaluations of different dimensions, forming an updated path state vector after multi-sub-channel interactive correction.

[0055] Based on the updated path state vector, multiple rounds of sub-channel evaluation and inter-channel interaction correction are repeatedly performed. In each round, the above steps are followed for sub-channel evaluation, state-sensitive factor extraction and broadcasting, and weighted correction of the path state vector. This continues until the changes in all sub-scores output by each sub-channel meet the preset convergence criteria. The preset convergence criteria can be set according to actual needs, such as sub-score changes being less than 1% or the trend of sub-score changes stabilizing over multiple rounds. Once the convergence criteria are met, a fusion process is performed, where the sub-scores output by each sub-channel are weighted and summed according to their importance to the overall path performance, outputting a comprehensive evaluation vector characterizing the candidate path under multi-dimensional constraints.

[0056] A multi-dimensional evaluation sub-channel comprehensively evaluates the initial candidate execution paths, considering their performance across multiple dimensions such as latency, load, and network transmission. Through the extraction and broadcasting of state-sensitive factors and the weighted correction of path state vectors, interaction and coupling between different evaluation dimensions are achieved, resulting in more accurate and comprehensive evaluation results. Furthermore, multi-round evaluation and convergence judgment mechanisms ensure the stability and reliability of the evaluation, thereby improving the effectiveness and accuracy of resource scheduling decisions.

[0057] Based on the comprehensive evaluation vector, perform path iterative search on N initial candidate execution paths to establish path iterative search results.

[0058] Furthermore, the path iterative search of N initial candidate execution paths based on the comprehensive evaluation vector also includes: establishing a search reference direction with respect to the N initial candidate execution paths based on the comprehensive evaluation vector; performing search optimization under the execution direction constraint after performing search perturbation on the search reference direction; and performing path iterative search of the execution path based on the search optimization result.

[0059] Furthermore, the search optimization under the execution direction constraint includes: mapping each candidate execution path to the corresponding search direction space according to the search reference direction, and configuring search priorities for each candidate execution path according to the comprehensive evaluation vector; after search perturbation, configuring multiple neighborhood perturbation paths for each candidate execution path, wherein the neighborhood perturbation paths are constructed by locally replacing or adjusting the node selection, task allocation ratio, or transmission link in the path; calculating the updated path state vector of each neighborhood perturbation path, performing effectiveness evaluation using the updated path state vector, and using the effectiveness evaluation results for search optimization.

[0060] Specifically, after obtaining the comprehensive evaluation vector corresponding to each candidate execution path, a search reference direction is constructed based on the comprehensive evaluation vector. The comprehensive evaluation vector is a multi-dimensional performance characterization result obtained by fusing latency sub-scores, load sub-scores, and network transmission sub-scores. The search reference direction refers to a directional vector used to indicate the path optimization trend, reflecting the dimensions that the current path should prioritize for improvement in the multi-dimensional constraint space. For example, when a path has a low latency sub-score but a good load and network score, its search reference direction will preferentially point to the direction of reducing latency occupancy. This can be achieved by normalizing the comprehensive evaluation vector and extracting the deviations between each dimension and the ideal target value to construct the search reference direction. The search reference direction provides clear guidance for the subsequent search process, ensuring that the search proceeds towards a better path.

[0061] After establishing the search reference direction, each candidate execution path is mapped to the corresponding search direction space. The search direction space is a directional search coordinate system constructed based on the multi-dimensional evaluation constraint space, with the search reference direction as the coordinate axis. Each candidate execution path has a corresponding search reference direction and its current position in the search direction space. The candidate paths are then sorted according to the comprehensive evaluation vector, and search priorities are configured. For example, paths with higher comprehensive scores and still room for optimization are prioritized for in-depth searching to improve overall search efficiency.

[0062] After performing search perturbation, multiple neighborhood perturbation paths are configured for each candidate execution path. Search perturbation aims to break the local optimum of the current path and introduce new possibilities. A neighborhood perturbation path refers to a new path generated through local adjustments while maintaining the basic structure of the original path. This can be achieved by locally replacing or adjusting nodes, task allocation ratios, or transmission links within the path. For example, in node selection, a node in the path can be randomly selected and replaced with a node with a lower load. In task allocation, the proportion of tasks undertaken by different nodes on the path can be adjusted. In transmission links, some transmission links in the path can be replaced or switched to transmission links with lower congestion levels. The data sources required for perturbation operations include node load monitoring data, link status monitoring data, and task execution configuration parameters from the resource scheduling system. For example, in a network path that originally passed through nodes A, B, and C, node B can now be replaced with the lower-loaded node D, constructing a new neighborhood perturbation path. By constructing multiple neighborhood perturbation paths, the search range can be expanded, increasing the chance of finding a better path.

[0063] For each disturbed path, its path state vector is recalculated, including latency, load, and network metrics at each stage, to obtain an updated path state vector. This updated path state vector is generated based on the original path state vector, taking into account changes in path characteristics after neighborhood disturbances. The updated path state vector is then input into a multi-dimensional evaluation sub-channel to obtain an updated comprehensive evaluation vector. A validity evaluation is then performed on the updated path state vector. This validity evaluation determines whether the disturbed path meets basic constraints, such as total latency not exceeding end-to-end latency constraints, node load not exceeding thresholds, and link reachability meeting requirements. The comprehensive evaluation vector is then used to determine whether the path offers performance improvement compared to the original path. Disturbed paths that meet the constraints and have a better comprehensive evaluation than the original path are retained and replace the original path in the next round of search. Paths that do not meet the constraints or exhibit performance degradation are pruned and eliminated.

[0064] Under directional constraints, the entire process from establishing the search reference direction to search optimization is repeated. The directional constraint means that path optimization must proceed along the search reference direction, prioritizing paths that improve the target dimension, such as reducing latency or load, thus avoiding ineffective or reverse searches. Through continuous iterative searching, the path set is gradually optimized, continuously improving the overall performance of the paths. When the change in the comprehensive evaluation vector falls below a preset threshold (e.g., less than 1%) in multiple consecutive iterations, or when the maximum number of iterations (100) is reached, the search process terminates, and the required path iteration search result is output.

[0065] By establishing a search reference direction and introducing neighborhood perturbations, the candidate paths are refined based on the multidimensional evaluation results, bringing the initial candidate paths closer to the optimal solution. At the same time, the search efficiency is improved by constraint pruning and priority control, avoiding blind traversal, and finally obtaining a path iteration search result with excellent comprehensive performance, thereby improving the accuracy and real-time performance of the overall scheduling decision.

[0066] Utilize path iteration search results to configure an elastic scheduling strategy, and execute intelligent scheduling management of target low-latency service requests according to the elastic scheduling strategy.

[0067] Furthermore, the elastic scheduling strategy is distributed to the corresponding cloud-side resource nodes and edge resource nodes to perform resource elastic scheduling management.

[0068] Specifically, after obtaining the path iteration search results, a scheduling strategy space is constructed based on the path iteration search results. Each path corresponds to a set of connection relationships and calculation processing order between cloud-side resource nodes, edge resource nodes, and terminal resource nodes. The resource types, processing stages, and transmission link relationships corresponding to each node in the path are structured and solidified to form a set of scheduling execution links consistent with the business execution process. The structured solidification refers to converting the node order relationship, resource attribute information, and link connection relationship in the path iteration search results into structured scheduling link data that the scheduling system can directly parse for subsequent resource scheduling instruction generation and execution control.

[0069] Then, the set of scheduling execution links is bound to the real-time monitored node computing power, load status, and network connectivity, so that each hop node in each path corresponds to a specific schedulable resource instance, thereby transforming the execution path from an optimization result into an executable resource mapping relationship. Based on this, the comprehensive evaluation vector obtained from path iterative search is used as a constraint reference and embedded into the scheduling control logic as a scheduling priority basis. This ensures that when executing low-latency service requests, tasks are flowed and processed according to the cloud-edge-device collaborative order determined by the execution path, and corresponding resource scheduling instruction sets are generated and sent to cloud-side and edge-side nodes. Each node completes the calculation and forwarding according to the predetermined path structure, thus forming an elastic scheduling strategy based on path iterative search results. This strategy guides the adaptive allocation and switching of cloud, edge, and device resources under different service load fluctuation conditions, achieving unified scheduling management of low-latency service requests.

[0070] After the elastic scheduling strategy is formulated, it is distributed to the corresponding cloud-side resource nodes and edge resource nodes through the resource management platform. The resource management platform has the functions of strategy distribution and resource monitoring. The resource management platform encapsulates the elastic scheduling strategy in a specific data format, such as JSON, and then sends the elastic scheduling strategy data to the corresponding cloud-side resource nodes and edge resource nodes through network communication protocols, such as HTTP. After receiving the elastic scheduling strategy data, the resource nodes parse and execute it, and perform elastic scheduling management of resources according to the strategy requirements. Among them, cloud-side resource nodes perform global computing scheduling according to the elastic scheduling strategy, edge resource nodes perform local task offloading and real-time processing, and terminal resource nodes dynamically adjust their data upload behavior based on feedback, thus forming a closed-loop scheduling system for cloud-edge-device collaboration.

[0071] By utilizing path iteration search results to configure elastic scheduling strategies and intelligently scheduling and managing target low-latency service requests according to the strategies, adaptive adjustment of cloud-edge-device resources is achieved. This ensures that low-latency services can still stably meet latency constraints in complex network and load fluctuation environments, improves the latency stability of low-latency service requests, the real-time performance and efficiency of scheduling decisions, and enhances the efficiency and resource utilization of cloud-edge-device collaborative execution.

[0072] Furthermore, during the generation process, the elastic scheduling strategy performs redundant backup configuration of the resource allocation scheme corresponding to the candidate execution path. When the performance degradation of the main path meets the predetermined conditions during execution, it switches to the alternative path. The switching of the alternative path is triggered based on the real-time monitored latency deviation.

[0073] Specifically, during the generation of elastic scheduling strategies, candidate execution paths are determined based on path iteration search results, and redundant backup configurations are simultaneously performed on the resource allocation schemes corresponding to these paths. Redundant backup configuration refers to preloading resource mapping relationships that are consistent with or highly similar to the main path structure into the scheduling system with low priority, without changing the main path's execution logic. This is used for rapid execution switchover in abnormal situations. The scheduling system binds and maps cloud-side, edge-side, and terminal-side nodes in the main path to their corresponding available alternative resource instances, and constructs a two-layer scheduling mapping table: one layer is the main execution mapping table for normal business request scheduling; the other layer is a backup mapping table for quickly taking over execution in case of failure or performance degradation, thereby achieving rapid reconfiguration of execution paths.

[0074] During operation, based on distributed link probes, node monitoring agents, and task execution logs, path execution status data is continuously collected, including end-to-end latency, node processing latency, link transmission latency, and queue queuing time. This data is then compared with the predicted latency baseline obtained during the path iterative search phase to calculate the latency deviation. The latency deviation is defined as the difference or relative deviation ratio between the actual execution latency and the predicted optimal latency. When the latency deviation exceeds a preset threshold, such as exceeding 5% or showing a continuous upward trend for more than three consecutive sampling periods, the main path is considered to have experienced performance degradation.

[0075] Once the triggering conditions are met, a path switching operation is performed based on the pre-built backup mapping table. This involves quickly redirecting the execution flow of the current business request from the main path mapping to the resource instance corresponding to the alternative path, and synchronously updating the task scheduling status and execution context information. This allows the business request to continue to complete the cloud-edge-device collaborative processing under the new path structure, thereby achieving the ability to quickly reroute the execution path and ensuring that low-latency services can continue to run stably under abnormal conditions.

[0076] By introducing a redundancy backup mechanism and a dynamic switching strategy based on latency deviation, the elastic scheduling strategy is equipped with runtime fault tolerance and adaptive recovery capabilities. This enables the stable execution of low-latency services even under conditions of cloud-edge-device resource load fluctuations or network performance degradation, thereby improving service continuity and reliability.

[0077] Example 2, based on the same inventive concept as the intelligent converged terminal cloud-edge resource elastic scheduling method for low-latency scenarios in the aforementioned examples, such as... Figure 2 As shown, this application provides an intelligent converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios, wherein the intelligent converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios includes: The constraint space construction module 11 is used to obtain the end-to-end latency constraints of the target low-latency service request, decompose the end-to-end latency constraints into terminal processing constraints, edge transmission constraints, and computation constraints, and construct a multi-dimensional evaluation constraint space corresponding to the decomposition results; the candidate path setting module 12 is used to set N initial candidate execution paths using the multi-dimensional evaluation constraint space; the evaluation analysis module 13 is used to map the N initial candidate execution paths into path state vectors, input them into the multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis, and output a comprehensive evaluation vector; the path search module 14 is used to perform path iterative search of the N initial candidate execution paths according to the comprehensive evaluation vector, and establish path iterative search results; the scheduling management module 15 is used to configure an elastic scheduling strategy using the path iterative search results, and perform intelligent scheduling management of the target low-latency service request according to the elastic scheduling strategy.

[0078] Furthermore, the multidimensional evaluation constraint space is used to characterize the constraint boundaries of different execution paths in terms of latency, load consumption, and network transmission overhead.

[0079] Furthermore, the candidate path setting module 12 is also used to: construct a convergence potential energy distribution model of the feasible region of the path based on the coupling relationship between the constraints in the multidimensional evaluation constraint space, wherein the convergence potential energy distribution model is used to evaluate the trend strength of different cloud, edge, and terminal combination execution paths in meeting the convergence of the target low-latency service request; calculate the potential energy gradient of the cloud-side execution path, edge-side execution path, and terminal-side execution path in the convergence potential energy distribution model to determine the convergence attraction strength of different path combinations in the feasible region; and generate N initial candidate execution paths with convergence guidance characteristics in the feasible region based on the convergence attraction strength.

[0080] Furthermore, the candidate path setting module 12 is also used to: divide cloud-side resource nodes, edge resource nodes, and terminal resource nodes into an ordered hierarchical structure according to the task processing stage, and construct an initial resource association graph containing node attributes and inter-node transmission overhead based on the computing power, current load status, and network connectivity of each node; map the convergence potential energy distribution model onto the initial resource association graph, assign the corresponding potential energy value to the nodes in the graph, and calculate the potential energy change of the edge weights between nodes according to the latency contribution, bandwidth occupancy, and computation queuing latency between nodes, forming a weighted path generation graph with potential energy gradient constraints; in the weighted path generation graph, Starting from terminal resource nodes and ending at cloud-side or edge resource nodes, a constrained path expansion is performed along the direction of decreasing potential energy to generate a candidate set of paths that meet convergence constraints. The number of effective surviving branches of the path expansion is counted, and an adaptive adjustment function for path size is constructed based on the decay rate of the number of effective paths with expansion depth. The number of paths corresponding to the convergence and stability interval of the adaptive adjustment function is used as the number N of candidate execution paths. The generated candidate path set is then subjected to multi-dimensional redundancy removal screening based on total path delay, resource occupancy balance, and network overhead, and path combinations that meet the difference threshold are retained as N initial candidate execution paths.

[0081] Furthermore, the evaluation and analysis module 13 is also used for: the multi-dimensional evaluation sub-channels include a delay sub-channel, a load sub-channel, and a transmission sub-channel; the delay sub-channel is used to output a delay sub-score based on the queuing delay and processing delay of each stage of the path; the load sub-channel constructs a resource consumption mapping relationship based on node computing power and real-time load status and outputs a resource consumption sub-score; the transmission sub-channel outputs a network transmission cost sub-score based on link bandwidth, congestion level, and transmission stability; after each sub-channel completes the calculation of all sub-scores for the current round, the corresponding state sensitivity factor is extracted, and the state sensitivity factor is used to characterize the sub-score pair. The gradient affecting the overall performance of the path is used to transmit sub-scores and state-sensitive factors to other sub-channels via an inter-channel broadcasting mechanism. Based on the state-sensitive factors, a weighted correction is performed on the corresponding feature components in the path state vector, so that the path state vector represents the coupling effect of cross-dimensional constraints in the next round of input, forming an updated path state vector after multi-sub-channel interactive correction. Based on the updated path state vector, multiple rounds of sub-channel evaluation and inter-channel interactive correction are repeated until the changes in all sub-scores output by each sub-channel meet the preset convergence conditions. After fusion processing, a comprehensive evaluation vector is output to characterize the overall performance of the candidate path.

[0082] Furthermore, the path search module 14 is also used to: establish a search reference direction with N initial candidate execution paths based on the comprehensive evaluation vector; perform search optimization under the execution direction constraint after performing search perturbation on the search reference direction; and perform path iterative search of the execution path based on the search optimization result.

[0083] Furthermore, the path search module 14 is also used to: map each candidate execution path to the corresponding search direction space according to the search reference direction, and configure the search priority for each candidate execution path according to the comprehensive evaluation vector; after search perturbation, configure multiple neighborhood perturbation paths for each candidate execution path, wherein the neighborhood perturbation paths are constructed by locally replacing or adjusting the node selection, task allocation ratio or transmission link in the path; calculate the updated path state vector of each neighborhood perturbation path, perform effectiveness evaluation using the updated path state vector, and perform search optimization using the effectiveness evaluation results.

[0084] Furthermore, the scheduling management module 15 is also used to: distribute the elastic scheduling strategy to the corresponding cloud-side resource nodes and edge resource nodes, and perform resource elastic scheduling management.

[0085] Furthermore, the scheduling management module 15 is also used to: during the generation process of the elastic scheduling strategy, perform redundant backup configuration of the resource allocation scheme corresponding to the candidate execution path, and switch to the alternative path when the performance degradation of the main path meets the predetermined conditions during execution, and the switching of the alternative path is triggered based on the real-time monitored latency deviation.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The method and specific examples of elastic scheduling of cloud-edge resources for intelligent converged terminals in low-latency scenarios described in Embodiment 1 are also applicable to the elastic scheduling system of cloud-edge resources for intelligent converged terminals in low-latency scenarios in this embodiment. Through the foregoing detailed description of the elastic scheduling method of cloud-edge resources for intelligent converged terminals in low-latency scenarios, those skilled in the art can clearly understand the elastic scheduling system of cloud-edge resources for intelligent converged terminals in low-latency scenarios in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0088] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios, characterized in that, The method includes: Obtain the end-to-end latency constraints of the target low-latency service request, decompose the end-to-end latency constraints into terminal processing constraints, edge transmission constraints and computation constraints, and construct a multi-dimensional evaluation constraint space corresponding to the decomposition results; N initial candidate execution paths are set using the multidimensional evaluation constraint space; After mapping the N initial candidate execution paths to path state vectors, the vectors are input into the multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation analysis and output a comprehensive evaluation vector. Based on the comprehensive evaluation vector, perform path iterative search on N initial candidate execution paths to establish path iterative search results; Utilize path iteration search results to configure an elastic scheduling strategy, and execute intelligent scheduling management of target low-latency service requests according to the elastic scheduling strategy.

2. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 1, characterized in that, The multidimensional evaluation constraint space is used to characterize the constraint boundaries of different execution paths in terms of latency, load consumption, and network transmission overhead.

3. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 2, characterized in that, N initial candidate execution paths are set using the multidimensional evaluation constraint space, including: Based on the coupling relationship between the constraints in the multidimensional evaluation constraint space, a convergence potential energy distribution model of the feasible path region is constructed. The convergence potential energy distribution model is used to evaluate the trend strength of different cloud, edge, and terminal combination execution paths to meet the convergence of target low-latency service requests. The potential energy gradient of the cloud-side execution path, edge-side execution path and terminal-side execution path is calculated in the convergence potential energy distribution model to determine the convergence attraction strength of different path combinations in the feasible region. Based on the convergence attraction strength, N initial candidate execution paths with convergence guidance characteristics are generated within the path feasible region.

4. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 3, characterized in that, Within the feasible path region, N initial candidate execution paths with convergence guidance properties are generated, which also includes: Cloud-side resource nodes, edge resource nodes, and terminal resource nodes are divided into an ordered hierarchical structure according to the task processing stage. Based on the computing power, current load status, and network connectivity of each node, an initial resource association graph containing node attributes and inter-node transmission overhead is constructed. The convergent potential energy distribution model is mapped to the initial resource association graph, and the nodes in the graph are assigned corresponding potential energy values. Based on the delay contribution, bandwidth occupancy and queuing delay between nodes, the potential energy change of the edge weights between nodes is calculated to form a weighted path generation graph with potential energy gradient constraints. In the weighted path generation graph, starting from the terminal resource node and ending at the cloud-side resource node or edge resource node, a constrained path expansion is performed along the direction of decreasing potential energy to generate a set of path candidate nodes that satisfy convergence constraints. The number of effective surviving branches of the path expansion is counted, and an adaptive adjustment function for path size is constructed based on the decay rate of the number of effective paths with the expansion depth. The number of paths corresponding to the convergence and stability interval of the adaptive adjustment function for path size is used as the number of candidate execution paths N. The generated path candidate set is filtered for redundancy removal in multiple dimensions based on total path latency, resource utilization balance, and network overhead. Path combinations that meet the difference threshold are retained as N initial candidate execution paths.

5. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 1, characterized in that, Inputting data into the multidimensional evaluation subchannel allows for multidimensional evaluation analysis, including: The multi-dimensional evaluation sub-channel includes a delay sub-channel, a load sub-channel, and a transmission sub-channel. The delay sub-channel is used to output a delay sub-score based on the queuing delay and processing delay at each stage of the path. The load sub-channel constructs a resource consumption mapping relationship based on the node's computing power and real-time load status and outputs a resource consumption sub-score. The transmission sub-channel outputs a network transmission cost sub-score based on link bandwidth, congestion level, and transmission stability. After each sub-channel completes the calculation of all sub-scores for the current round, the corresponding state sensitivity factor is extracted. The state sensitivity factor is used to characterize the gradient of the impact of the sub-score on the overall performance of the path. The sub-scores and state sensitivity factors are then transmitted to the other sub-channels through the inter-channel broadcasting mechanism. Based on the state sensitivity factor, a weighted correction process is performed on the corresponding feature components in the path state vector so that the path state vector represents the coupling effect of cross-dimensional constraints in the next round of input, forming an updated path state vector after multi-sub-channel interactive correction. Based on the updated path state vector, multiple rounds of sub-channel evaluation and inter-channel interaction correction are repeatedly performed until the changes in all sub-scores output by each sub-channel meet the preset convergence conditions. After fusion processing, a comprehensive evaluation vector is output to characterize the overall performance of the candidate path.

6. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 1, characterized in that, The path iterative search based on the comprehensive evaluation vector, performing N initial candidate execution paths, further includes: A search reference direction is established based on the comprehensive evaluation vector and the N initial candidate execution paths; After performing a search perturbation on the search reference direction, perform search optimization under directional constraints; Execute path iterative search based on the search optimization results.

7. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 6, characterized in that, Search optimization under execution direction constraints includes: Based on the search reference direction, each candidate execution path is mapped to the corresponding search direction space, and the search priority is configured for each candidate execution path according to the comprehensive evaluation vector; After the search perturbation, multiple neighborhood perturbation paths are configured for each candidate execution path. The neighborhood perturbation paths are constructed by locally replacing or adjusting the node selection, task allocation ratio, or transmission links in the path. Calculate the updated path state vector for each neighborhood disturbance path, perform an effectiveness evaluation using the updated path state vector, and use the effectiveness evaluation results for search optimization.

8. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 1, characterized in that, The elastic scheduling strategy is distributed to the corresponding cloud-side resource nodes and edge resource nodes to perform resource elastic scheduling management.

9. The method for elastic scheduling of cloud-edge resources in intelligent converged terminals for low-latency scenarios as described in claim 1, characterized in that, During the generation process, the elastic scheduling strategy performs redundant backup configuration of the resource allocation scheme corresponding to the candidate execution path. When the performance degradation of the main path meets the predetermined conditions during execution, it switches to the alternative path. The switching of the alternative path is triggered based on the real-time monitored latency deviation.

10. A smart converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios, characterized in that, The steps for implementing the intelligent converged terminal cloud-edge resource elastic scheduling method for low-latency scenarios according to any one of claims 1 to 9, wherein the intelligent converged terminal cloud-edge resource elastic scheduling system for low-latency scenarios comprises: The constraint space construction module is used to obtain the end-to-end latency constraints of the target low-latency service request, decompose the end-to-end latency constraints into terminal processing constraints, edge transmission constraints and computation constraints, and construct a multi-dimensional evaluation constraint space corresponding to the decomposition results. The candidate path setting module is used to set N initial candidate execution paths using the multidimensional evaluation constraint space; The evaluation and analysis module is used to map N initial candidate execution paths into path state vectors, input them into the multi-dimensional evaluation sub-channel to perform multi-dimensional evaluation and analysis, and output a comprehensive evaluation vector. The path search module is used to perform path iterative search on N initial candidate execution paths based on the comprehensive evaluation vector, and to establish path iterative search results. The scheduling management module is used to configure elastic scheduling strategies using path iteration search results, and to perform intelligent scheduling management of target low-latency service requests according to the elastic scheduling strategies.

Images