A task scheduling method based on dynamic cloud edge collaboration

Abstract

The present application belongs to the technical field of mobile edge computing, and discloses a task scheduling method based on dynamic cloud edge cooperation. Through link level modeling and graph neural network feature extraction, the present application realizes adaptive expansion of different numbers of terminals and servers; in combination with the pruning update of the near-end strategy optimization algorithm and the generalized advantage estimation method, the present application effectively avoids instability and overestimation problems in training. The present application can not only fully utilize the elastic computing power of the cloud and the low latency characteristics of the edge, but also adapt to the dynamically changing network size and terminal demand, and can realize flexible task offloading and dynamic resource allocation, effectively reduce task processing delay in a dynamic environment, improve resource utilization, and maintain good scalability and stability under the condition of terminal quantity and server scale fluctuation, solve the computing resource bottleneck and task scheduling rigidity problems existing in the virtual reality video service of the existing mobile edge computing, and significantly improve the user's interactive experience.

Description

Technical Field

[0001] This invention relates to the field of mobile edge computing technology, and in particular to a task scheduling method based on dynamic cloud-edge collaboration. Background Technology

[0002] In recent years, with the rapid development of virtual reality (VR) technology, global VR device shipments have continued to grow. According to statistics from International Data Corporation (IDC), VR device shipments increased by 12.8% year-on-year in the third quarter of 2024. At the same time, more and more end users are using head-mounted displays with resolutions reaching 4K or even 8K and refresh rates up to 120Hz. While these high-performance devices significantly enhance the immersive experience, their demand for computing and network resources is also growing exponentially.

[0003] Mobile edge computing (MEC) technology, by deploying computing resources at the network edge, can shorten task transmission paths, reduce latency, and meet the real-time and bandwidth-sensitive needs of applications such as VR, AR, and cloud gaming. However, because VR applications involve complex computing tasks such as high-definition video encoding and decoding, real-time scene rendering, and user interaction synchronization, relying solely on edge servers often faces insufficient computing power, especially in high-concurrency, multi-user scenarios where resource congestion and service quality degradation are more likely to occur.

[0004] Most existing task scheduling methods are based on static assumptions, such as assuming that the number of terminal devices and server resources remain constant. This approach has significant shortcomings in real-world dynamic network environments: when the number of terminals suddenly increases or network link conditions fluctuate, static scheduling can easily lead to task queue overflow, a sharp increase in average processing latency, and even cause severe dizziness and discomfort for users, seriously affecting the VR experience. Furthermore, static models lack adaptability to dynamic resource changes, making it difficult to promote and apply them in edge-cloud collaborative systems of different scales.

[0005] Existing technologies attempt to combine intelligent optimization methods such as Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNN) to improve the flexibility of edge-cloud task scheduling. Traditional intelligent optimization methods, such as simulated annealing and genetic algorithms, are not suitable for systems with millisecond-level latency sensitivity. Due to the lack of historical data required to support the training of supervised learning models, reinforcement learning (RL) methods with environmental interaction capabilities have become an ideal alternative.

[0006] Reinforcement learning (RL) is a machine learning method that learns decision-making policies through interaction with the environment. Its core principle is that an agent chooses an action given a state, and the environment responds with a new state and a corresponding reward. The agent iteratively adjusts its policy based on this feedback to maximize long-term cumulative rewards. In RL algorithms, Proximal Policy Optimization (PPO) is widely used due to its stability and effectiveness. PPO optimizes the policy network by limiting the size of policy updates during each iteration, thus maintaining training stability and effectively overcoming the dimensionality curse associated with the discrete action space in traditional Deep Q-Network (DQN) algorithms. The importance sampling mechanism in PPO significantly improves sampling efficiency, which is particularly important in cloud-edge collaborative environments requiring real-time responses to millisecond-level VR interactions.

[0007] When tasks involve graph-structured data with complex dependencies, Multilayer Perceptrons (MLPs) often fail to capture the inter-node dependencies and topological structure. In such cases, Graph Neural Networks (GNNs) offer a significant advantage. By leveraging message passing mechanisms, GNNs explicitly model the connections between nodes and extract structural information feature representations. Therefore, GNNs exhibit superior performance in scenarios such as task offloading and resource allocation.

[0008] Currently, existing methods have proposed reducing cold-start overhead through deep reinforcement learning or optimizing content caching based on graph neural networks. However, most of these methods still remain at the node-level feature modeling level, lacking fine-grained characterization of dynamic relationships at the link level, making it difficult to fully cope with complex and ever-changing VR scenarios. Existing technologies have limitations in areas such as static resource scheduling, insufficient system scalability, and excessively high task latency. Therefore, there is an urgent need for a task scheduling method that can dynamically coordinate between the cloud and the edge. This method can fully utilize the elastic computing power of the cloud and the low latency characteristics of the edge, while also adapting to dynamically changing network scale and terminal requirements, in order to ensure the real-time performance and stability of VR video services. Summary of the Invention

[0009] Existing methods, which are mostly based on static assumptions and cannot adapt to dynamic fluctuations in the number of terminals and server resources, have limitations in areas such as static resource scheduling, insufficient system scalability, and excessively high task latency. This invention provides a task scheduling method based on dynamic cloud-edge collaboration, aiming to achieve flexible task offloading and dynamic resource allocation, ensure the real-time performance and stability of VR video services, solve the computing resource bottlenecks and rigid task scheduling problems existing in existing mobile edge computing (MEC) in virtual reality (VR) video services, and improve the user interaction experience.

[0010] To achieve the above objectives, this invention provides a task scheduling method based on dynamic cloud-edge collaboration, comprising the following steps:

[0011] Step 1: System Modeling and Heterogeneous Graph Construction:

[0012] Step 1.1: The server includes a cloud server V. c Edge Server V s Establish a cloud server V c Edge Server V s and terminal V u The cloud-edge-device system model, under the cloud-edge collaborative architecture, models cloud server nodes, edge server nodes, and terminal nodes as a heterogeneous graph, where the heterogeneous graph G = (V u V s V c , E), where E represents all possible terminal-server connections;

[0013] Terminal node characteristics include task data volume d i Edge server node characteristics include computing power f j The link characteristics between the terminal node and the server node include the channel gain h. i，j ;

[0014] Step 1.2: Divide each terminal task into several sub-tasks according to the proportion, and offload them to different servers to establish transmission constraints and computing resource constraints;

[0015] Step 1.3: Establish a total task latency model, including transmission latency and computation latency, and store it in the initial scheduling table;

[0016] Step 2: Generate scheduling strategy based on graph neural network (GNN):

[0017] Step 2.1: Heterogeneous graph feature extraction: Graph neural network (GNN) is used to aggregate and update terminal node features, edge server node features, and link features between terminal nodes and server nodes, and to extract the relationships between task allocation, power allocation, and resource allocation.

[0018] Step 2.2: Scheduling Policy Matrix Output: Through the forward output of the Graph Neural Network (GNN), three types of scheduling policy matrices are generated, including the task allocation matrix A. task Power allocation matrix A power and resource allocation matrix A comp Task allocation matrix A task Including the offloading ratio between terminals and servers, power allocation matrix A power Including link transmission power allocation, resource allocation matrix A comp This includes the allocation ratio of server computing resources;

[0019] Step 2.3: Reward signal definition: Minimize the total delay as the optimization objective, and obtain the reward function after normalization;

[0020] Step 3: Update the scheduling policy using the Proximity-to-End (PPO) optimization algorithm:

[0021] Step 3.1: Model the scheduling problem as a Markov Decision Process (MDP). The state s includes the task data volume, channel gain, and computing power. The scheduling policy matrix a is a matrix of three types of scheduling policies, and the reward function r is... e This is the normalized time delay function;

[0022] Step 3.2: Use the objective function L clip =min(r t ×A t ,clip(r t ,1-θ,1+θ)×A t To achieve stable iterative updates of the scheduling strategy, where r t A represents the ratio of the probability of the new scheduling policy and the old scheduling policy being in the same state. t The dominant function is represented by θ, the clipping parameter is represented by θ, and the clipping loss function is clip(r). t (1-θ, 1+θ) represents r t The range is limited to [1-θ, 1+θ]. min() represents the minimum value operation.

[0023] Step 3.3: Calculate the dominance value using the generalized dominance estimation method (GAE) to avoid scheduling policy collapse and improve convergence speed;

[0024] Step 3.4: In each iteration, the scheduling policy matrix generated by the graph neural network (GNN) and the reward signal from step 2.3 are used together for optimization, ultimately resulting in a dynamically adjusted policy network;

[0025] Step 4: Dynamic cloud-edge collaborative execution:

[0026] Step 4.1: When edge server resources are sufficient, tasks should be executed on the local edge server first;

[0027] Step 4.2: When the edge server is overloaded, some subtasks can be offloaded to the cloud server;

[0028] Step 4.3: When there are multiple neighboring edge servers, tasks can be distributed through edge server-to-edge server collaboration to reduce the risk of single-point overload.

[0029] Furthermore, in step 1.2, the connection between terminals i and j is represented by link (i,j), and the task allocation ratio is defined as x. i,jThe transmission power is p i,j The computing resources allocated to the server are f i,j Then the transmission constraints must satisfy x i,j ≥0, where ∑ represents the summation operation, M represents the number of terminals, and x i，cloud This represents the task allocation ratio between terminal i and the cloud server;

[0030] Computational resource constraints must be met Among them, f i，j This represents the computing resources allocated to the server. Representing cloud computing resources, f i This represents the total resources of the server, ensuring that the total resources of the server do not exceed its maximum computing capacity.

[0031] Furthermore, step 1.3 establishes a total task latency model, including transmission latency and computation latency, and stores it in the initial scheduling table, including the following steps:

[0032] Step 1.3.1: Express the transmission delay as... in, Let link (i, j) represent the connection between terminals i and j, and x i,j d represents the task allocation ratio between terminals i and j. i For the amount of task data, r i，j Let b be the link transmission rate between terminals i and j, b be the bandwidth, and p be the bandwidth. i，j Let σ be the transmission power between terminals i and j. 2 h is the noise power. i，j p is the channel gain between terminals i and j. k，j h represents the transmission power between terminals k and j. k，j Let N be the channel gain between terminals k and j, and N be the number of terminals.

[0033] Step 1.3.2: Express the calculation delay as... Where c is the computational cost constant required per unit of data, and f i，j Computing resources allocated to the server;

[0034] Step 1.3.3: Total delay is expressed as The total delay value serves as the core objective function for subsequent scheduling optimization and is also stored in the initial scheduling table.

[0035] Furthermore, the reward function in step 2.3 is expressed as follows: The reward value is in the range [0,1].

[0036] T smo =(1-s f )×Ti,j +s f ×T i,j

[0037]

[0038] middle, T smo To smooth out the delay, These represent the maximum and minimum values ​​over time, dynamically updated using an exponential moving average, s. f The value of is 0.2, which is the time smoothing factor, and the value of a′ is 0.1, which is the exponential moving average dynamic update coefficient. and They represent time T respectively i，j The minimum and maximum values ​​of dynamically updated task processing latency.

[0039] Beneficial Effects: This invention provides a task scheduling method based on dynamic cloud-edge collaboration. Through link-level modeling and graph neural network (GNN) feature extraction, it achieves adaptive scaling for different numbers of terminals and servers, overcoming the limitation of fixed input and output dimensions in traditional multilayer perceptron (MLP) methods. By combining the pruning and updating of the proximal policy optimization algorithm (PPO) and the generalized advantage estimation (GAE) method, it effectively avoids instability and overestimation problems during training, ensuring efficient and stable task scheduling. Experimental results show that the proposed dynamic edge-cloud collaborative scheduling method (DECS) reduces task processing latency by 8.52% and 20.54% compared to pure edge computing and pure cloud computing architectures, respectively, significantly improving the real-time interactive experience in VR services. Attached Figure Description

[0040] Figure 1 This is a basic structural diagram of the cloud-edge collaboration scenario involved in the embodiments of the present invention;

[0041] Figure 2 This is an algorithm structure diagram of the scalable dynamic edge-cloud collaborative scheduling method DECS in the cloud-edge collaborative scenario involved in the embodiments of the present invention;

[0042] Figure 3 This is a comparison chart of rewards under different models involved in the embodiments of the present invention;

[0043] Figure 4 This is a schematic diagram illustrating the task processing latency of different models under different numbers of servers in embodiments of the present invention;

[0044] Figure 5 This is a schematic diagram illustrating task processing latency for different numbers of edge servers involved in embodiments of the present invention;

[0045] Figure 6This is a schematic diagram illustrating the task processing latency for different numbers of terminals involved in embodiments of the present invention. Detailed Implementation

[0046] like Figures 1 to 6 As shown, to address the problems of rigid resource allocation and inability to adapt to dynamic terminal scale and computing resource fluctuations in existing edge computing for virtual reality (VR) task scheduling, this invention provides a task scheduling method based on dynamic cloud-edge collaboration. The Dynamic Edge-Cloud Collaborative Scheduling (DECS) method is based on graph neural networks (GNNs) and near-end policy optimization (PPO). This method, based on a cloud-edge collaborative architecture and combined with a reinforcement learning policy update mechanism, can achieve flexible task offloading and optimized resource allocation under any number of terminals and edge nodes, thereby effectively reducing task processing latency for VR video services and improving the user interaction experience.

[0047] Example 1: As Figure 1 As shown, the system model of this invention consists of end users, edge servers, and cloud servers. End users are connected to edge servers via base stations, edge servers collaborate with each other via local area networks, and edge servers are connected to the cloud center via backhaul links. The specific implementation process is as follows: Figure 2 As shown.

[0048] Step 1: System Modeling and Task Graph Construction

[0049] In the initial task scheduling phase, the computing tasks generated by end users have not yet been assigned, and all tasks are processed through the cloud center by default. At this point, as the process begins, the task allocation relationship between the terminal and the server has not yet been established. Therefore, a system model and task graph must first be constructed for subsequent dynamic scheduling optimization, such as... Figure 3 As shown.

[0050] Step 1.1: Initialize nodes and links, and denote the terminal as V. u Edge server V s Cloud servers are denoted as V. c For terminal nodes, the characteristic is the amount of task data d. i For edge server nodes, the characteristic is computing power f j For cloud server nodes, the computing power is set to be significantly higher than that of the edge nodes, and the transmission link latency is set to be twice that of the edge links. Link characteristics include channel gain h. i,j This is used to represent the channel quality between the terminal and the server. Therefore, a heterogeneous graph G = (V) is constructed. u V s V c ,E), where E represents all possible terminal-server connections.

[0051] Step 1.2: Divide each terminal task into several sub-tasks according to the proportion, and offload them to different servers to establish transmission constraints and computing resource constraints;

[0052] Let link (i,j) represent the connection between terminals i and j, and define the task allocation ratio as x. i,j The transmission power is p i,j The computing resources allocated to the server are f i，j Then the transmission constraints must satisfy x i，j ≥0, where ∑ represents the summation operation, M represents the number of terminals, and x i，cloud This represents the task allocation ratio between terminal i and the cloud server;

[0053] Computational resource constraints must be met Among them, f i,j This represents the computing resources allocated to the server. Representing cloud computing resources, f i This represents the total resources of the server, ensuring that the total resources of the server do not exceed its maximum computing capacity.

[0054] Step 1.3: Establish a total task latency model, including transmission latency and computation latency, and store it in the initial scheduling table;

[0055] Includes the following steps:

[0056] Step 1.3.1: Express the transmission delay as... in, Let link (i,j) represent the connection between terminals i and j, and x i,j d represents the task allocation ratio between terminals i and j. i For the amount of task data, r i,j Let b be the link transmission rate between terminals i and j, b be the bandwidth, and p be the bandwidth. i,j Let σ be the transmission power between terminals i and j. 2 h is the noise power. i，j p is the channel gain between terminals i and j. k,j h represents the transmission power between terminals k and j. k,j Let N be the channel gain between terminals k and j, and N be the number of terminals.

[0057] Step 1.3.2: Express the calculation delay as... Where c is the computational cost constant required per unit of data, and f i,j Computing resources allocated to the server;

[0058] Step 1.3.3: Total delay is expressed as The total latency value serves as the core objective function for subsequent scheduling optimization and is stored in the initial scheduling table for later GNN modeling and PPO policy training. Through the initialization in step 1, the system completes the definition of terminal nodes, server nodes, and their link relationships during the modeling phase, and establishes the task splitting ratio, resource constraints, and latency calculation model, laying the foundation for subsequent feature extraction and scheduling optimization based on graph neural networks.

[0059] Step 2: Generate scheduling strategy based on graph neural network (GNN):

[0060] Step 2.1: Heterogeneous graph feature extraction: Graph neural network (GNN) is used to aggregate and update terminal node features, edge server node features, and link features between terminal nodes and server nodes, and to extract the relationships between task allocation, power allocation, and resource allocation.

[0061] Step 2.2: Scheduling Policy Matrix Output: Through the forward output of the Graph Neural Network (GNN), three types of scheduling policy matrices are generated, including the task allocation matrix A. task Power allocation matrix A power and resource allocation matrix A comp Task allocation matrix A task Including the offloading ratio between terminals and servers, power allocation matrix A power Including link transmission power allocation, resource allocation matrix A comp This includes the allocation ratio of server computing resources;

[0062] Step 2.3: Reward signal definition: Minimize the total delay as the optimization objective, and obtain the reward function after normalization;

[0063] To guide scheduling optimization, the goal of minimizing latency is transformed into a reward function, expressed as follows:

[0064] T smo =(1-s f )×T i,j +s f ×T i,j

[0065]

[0066] in, T smo To smooth out the delay, These represent the maximum and minimum values ​​over time, dynamically updated using an exponential moving average, s. f The value of is 0.2, which is the time smoothing factor, and the value of a′ is 0.1, which is the exponential moving average dynamic update coefficient. and They represent time T respectively i,j The minimum and maximum values ​​of dynamically updated task processing latency.

[0067] Step 3: Update the scheduling policy using the Proximity-to-End (PPO) optimization algorithm:

[0068] Step 3.1: Model the scheduling problem as a Markov Decision Process (MDP). The state s includes the task data volume, channel gain, and computing power. The scheduling policy matrix a is a matrix of three types of scheduling policies, and the reward function r is... e This is the normalized time delay function;

[0069] Step 3.2: Scheduling strategy update: The objective function used is represented as L clip =min(r t ×A t ,clip(r t ,1-θ,1+θ)×A t To achieve stable iterative updates of the scheduling policy, where r t A represents the ratio of the probability of the new scheduling policy and the old scheduling policy being in the same state. t The dominant function is represented by θ, the clipping parameter is represented by θ, and the clipping loss function is clip(r). t (1-θ, 1+θ) represents r t The objective function is limited to the range [1-θ, 1+θ], and min() represents the minimum value operation. The objective function combines the advantage function and the pruning loss function to achieve stable iterative updates of the scheduling strategy.

[0070] Step 3.3: Calculate the dominance value using the Generalized Dominance Estimation (GAE) method, denoted as A. t =δ t +γλA t+1 Where γ is the discount factor, λ is the factor for balancing bias and variance, and δ t For timing difference error, if δ t >0 indicates that the current action is better than expected; if δ t <0 indicates that the situation is worse than expected, which helps to prevent the scheduling strategy from collapsing and improve the convergence speed.

[0071] Step 3.4: In each iteration, the scheduling policy matrix generated by the graph neural network (GNN) and the reward signal from step 2.3 are used together for optimization, ultimately resulting in a dynamically adjusted policy network;

[0072] Step 4: Dynamic cloud-edge collaborative execution:

[0073] Step 4.1: When edge server resources are sufficient, tasks should be executed on the local edge server first;

[0074] Step 4.2: When the edge server is overloaded, some subtasks can be offloaded to the cloud server;

[0075] Step 4.3: When there are multiple neighboring edge servers, tasks can be distributed through edge server-to-edge server collaboration to reduce the risk of single-point overload.

[0076] This process ensures efficient task allocation and minimizes latency in dynamic environments.

[0077] To verify the effectiveness of this invention, the proposed Dynamic Edge-Cloud Collaborative Scheduling (DECS) method for dynamic cloud-edge collaborative task scheduling is compared with traditional edge computing and cloud computing methods. The comparison methods include:

[0078] 1) Pure edge computing architecture: All tasks are processed by edge servers;

[0079] 2) Pure cloud computing architecture: All tasks are offloaded to cloud servers for processing;

[0080] 3) Traditional MLP+DQN scheduling method: Uses a neural network with fixed input and output dimensions for task allocation.

[0081] In the experimental scenario, the number of terminals was set to twice the number of servers, and the performance of different methods in terms of training convergence, task processing latency, and resource utilization was compared through simulation.

[0082] like Figure 3 As shown, the reward value changes for different methods as the number of training iterations increases. It can be seen that the traditional MLP+DQN method experiences significant oscillations in the initial stages, has a slower convergence speed, and ultimately yields a lower reward value than the DECS method proposed in this invention. In contrast, the DECS method gradually stabilizes after 300 iterations, with the reward value approaching the optimal level, indicating that this invention is more stable in terms of scheduling policy learning.

[0083] like Figure 4 and Figure 5 As shown, the task processing latency of different methods is compared under different server scales. The results show that: when the number of servers is limited, the average task processing latency of the pure edge computing architecture is relatively high due to insufficient single-point computing power; although the pure cloud computing architecture has strong computing power, the additional latency caused by remote link transmission significantly increases the total latency; the DECS method proposed in this invention, under the cloud-edge collaborative architecture, can adaptively adjust the offloading ratio and resource allocation, and always maintain a low processing latency. Compared with the pure edge computing and pure cloud computing methods, the average task processing latency is reduced by 8.52% and 20.54%, respectively.

[0084] like Figure 6As shown, the latency trends of different methods under a surge in terminals are compared. When the number of terminals increases rapidly, the latency of the pure edge computing method increases linearly, while the performance of the pure cloud computing method degrades due to increased link load. The DECS method of this invention, by dynamically adjusting the cloud-edge offloading ratio, significantly slows down the latency growth curve and remains superior to the comparative methods in large-scale scenarios.

[0085] As shown in Table 1, the comparison results of CPU resource utilization show that, under the same scale, pure edge computing has excessively high utilization due to the limited single-machine capability, while pure cloud computing has idle resources. However, the DECS method of this invention can reasonably allocate the load between the cloud and the edge, so that the edge side utilization is reduced to about 68% and the cloud side utilization is stabilized at about 30%, thereby effectively avoiding the problems of single-point overload and resource waste.

[0086] Table 1 Comparison of CPU resource utilization results

[0087]

[0088] This invention provides a task scheduling method based on dynamic cloud-edge collaboration, shifting resource allocation from node-level features to link-level features. Each link between the terminal and the server is assigned an independent feature representation. Because the model processes features at the link level, rather than producing fixed-dimensional outputs at the node level, additional terminals and servers can be naturally accommodated by increasing the number of links without modifying the model's structure or parameters. This invention fully utilizes the elastic computing power of the cloud and the low latency of the edge, adapting to dynamically changing network scale and terminal demands. It enables flexible task offloading and dynamic resource allocation, effectively reducing task processing latency and improving resource utilization in dynamic environments. It maintains good scalability and stability even with fluctuations in the number of terminals and server scale. This solves the computational resource bottlenecks and rigid task scheduling problems existing in current mobile edge computing (MEC) for virtual reality (VR) video services, significantly improving the user's interactive experience.

[0089] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. However, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A task scheduling method based on dynamic cloud-edge collaboration, characterized in that, Comprising the following steps: Step 1: System modeling and heterogeneous graph construction: Step 1.1: The server includes a cloud server V. c Edge Server V s Establish a cloud server V c Edge Server V s and terminal V u The cloud-edge-device system model, under the cloud-edge collaborative architecture, models cloud server nodes, edge server nodes, and terminal nodes as a heterogeneous graph, where the heterogeneous graph G = (V u V s V c , E), where E represents all possible terminal-server connections; Terminal node characteristics include task data volume d i Edge server node characteristics include computing power f j Link characteristics between terminal node-server node include channel gain h i,j ; Step 1.2: Split each terminal task into several subtasks in proportion, respectively unload to different servers, establish transmission constraints and computing resource constraints; Step 1.3: Establish a total task delay model, including transmission delay and computing delay, and store it in the initial scheduling table; Step 2: Generate scheduling strategy based on graph neural network GNN: Step 2.1: Heterogeneous graph feature extraction: use graph neural network GNN to aggregate and update terminal node features, edge server node features, and link features between terminal node-server node, extract task allocation, power allocation, and resource allocation relationships; Step 2.2: Scheduling strategy matrix output: generate three types of scheduling strategy matrices including task allocation matrix A task , power allocation matrix A power and resource allocation matrix A comp through the forward output of the graph neural network GNN, wherein the task allocation matrix A task includes the offloading ratio between the terminal and the server, the power allocation matrix A power includes the link transmission power allocation, and the resource allocation matrix A comp includes the server computing resource allocation ratio; Step 2.3: Reward signal definition: minimize the total delay as the optimization objective, and obtain the reward function after normalization processing; Step 3: Update the scheduling strategy using the proximal policy optimization algorithm PPO: Step 3.1: Model the scheduling problem as a Markov Decision Process (MDP), the state s includes the task data volume, channel gain, computing capacity, the scheduling policy matrix a is the three kinds of scheduling policy matrix, the reward function r e is the normalized latency function; Step 3.2: Implementing the stable iterative update of the scheduling policy using the objective function L clip = min(r t , 1) * A t , clip(r t , 1-0, 1+0) * A t ) where r t represents the ratio of the probability of the new scheduling policy and the old scheduling policy in the same state, A t represents the advantage function, 0 represents the clipping parameter, the clipping loss function clip(r t , 1-0, 1+0) represents limiting r t in the range of [1-0, 1+0], and min() represents the minimum value operation. Step 3.3: Calculate the advantage value through the generalized advantage estimation method GAE to avoid scheduling strategy collapse and improve convergence speed; Step 3.4: In each iteration, the scheduling strategy matrix generated by the graph neural network GNN and the reward signal in step 2.3 participate in optimization together, and finally obtain the dynamic adjustment strategy network; Step 4: Dynamic cloud-edge collaborative execution: Step 4.1: When the edge server resources are sufficient, the task is preferentially executed on the local edge server; Step 4.2: When the edge server load is too high, part of the subtasks can be unloaded to the cloud server; Step 4.3: When there are multiple neighbor edge servers, the task can be allocated through edge server-edge server collaboration to reduce the risk of single point overload. 2.The method of claim 1, wherein, In step 1.2, the connection between terminals i and j is represented by link (i, j), and the task allocation ratio is defined as x i,j , the transmission power is p i,j , and the computing resources allocated to the server are f i,j The transmission constraint needs to be satisfied x i，j ≥ 0, where ∑ represents the summation operation, M represents the number of terminals, and x i，cloud represents the task allocation ratio between terminal i and the cloud server. Computing resource constraints need to be satisfied where f i，j represents the computing resource allocated to the server, represents the cloud computing resource, f i represents the total resource of the server, ensuring that the total resource of the server does not exceed its maximum computing capacity. 3.The method of claim 1, wherein, Step 1.3 establishes a total task delay model, including transmission delay and computing delay, and stores it in the initial scheduling table, comprising the following steps: Step 1.3.1: Express the transmission delay as where, Let link (i, j) denote the connection between terminals i and j, x i，j is the task allocation ratio between terminals i and j, d i is the task data volume, r i，j is the link transmission rate between terminals i and j, b is the bandwidth, p i，j is the transmission power between terminals i and j, σ 2 is the noise power, h i，j is the channel gain between terminals i and j, p k，j is the transmission power between terminals k and j, h k，j is the channel gain between terminals k and j, N is the number of terminals; Step 1.3.2: Express the computation latency as where c is a constant computation overhead constant for unit data, f i,j is the computation resource allocated to the server; Step 1.3.3: The total latency is expressed as The value of the total latency is stored in the initial schedule table as the core objective function for subsequent scheduling optimization. 4.The method of claim 1, wherein, The reward function in Step 2.3 is represented as The reward value is in the interval [0, 1], where, T smo = (1 - s f ) x T i，j + s f x T i，j T smo is the smoothed time delay, respectively represent the maximum and minimum values of time, which are dynamically updated by exponential moving average, s f is 0.2, and a' is 0.1, which are the time smoothing factor and the exponential moving average dynamic update coefficient, and respectively represent the dynamically updated minimum and maximum values of the task processing delay at time T i,j .

Images