Distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical cooperation

Abstract

The invention provides a distributed edge computing task scheduling method based on a dynamic hypergraph and hierarchical cooperation, and relates to the field of distributed edge computing. The method comprises the following steps: a gateway node obtains a task set, constructs a dynamic hypergraph model and combines with other system real-time indexes to generate a system state; the gateway node generates a task division decision through a reinforcement learning model based on a system state; the scheduler node carries out multi-agent cooperation decision making according to a task division decision in combination with local observation information, and task subsets are dispatched to a server in a distributed mode; the server executes the task and calculates the throughput rate and the energy consumption to generate a delay feedback signal; and based on the delayed feedback signal, optimizing the reinforcement learning model, and dynamically adjusting a task division decision. According to the method, the problems of task division and resource scheduling separation, insufficient high-order task dependency relationship modeling and lack of real-time coordination of hierarchical decision in the prior art are solved, and the energy consumption is reduced while the system throughput is ensured.

Description

Technical Field

[0001] This invention relates to the field of distributed edge computing, and in particular to a distributed edge computing task scheduling method based on dynamic hypergraphs and hierarchical collaboration. Background Technology

[0002] With the rapid development of artificial intelligence services, business demands are placing higher requirements on the real-time response capabilities of computing, and computing tasks are migrating from traditional centralized cloud computing to the network edge. Against this backdrop, distributed edge computing (DEC) has emerged as a key technological paradigm that integrates edge computing and parallel computing.

[0003] However, current distributed edge computing scheduling technologies have obvious bottlenecks: (1) Existing technologies usually separate the collaboration between "task partitioning" and "resource scheduling". Most studies are limited to optimizing the local scheduling within the downstream edge resource pool, and wrongly assume that the task partitioning of the upstream gateway is reasonable. As a result, even if the downstream scheduling is optimized well, the global energy efficiency limit is still limited by the improper partitioning of the upper layer. (2) Existing technologies usually use directed acyclic graphs (DAGs) to describe tasks. This method can only reflect the simple order between tasks, but ignores the implicit relationship between tasks. As a result, the system often incorrectly splits a group of tasks that should be processed together into different computing nodes, which leads to resource waste. (3) In the existing architecture, the gateway's traffic splitting decision and the scheduler's execution scheduling are usually carried out independently. The gateway lacks real-time perception of the actual resource congestion level or execution feedback inside the scheduler. This one-way open-loop decision mode leads the gateway to often make incorrect traffic splitting decisions that do not match the actual resource status, which leads to resource contention and performance degradation.

[0004] Therefore, there is an urgent need for a joint optimization method that can integrate task partitioning and resource scheduling to break the global energy efficiency limitations caused by improper upper-level distribution, perceive the actual status of resources in real time, and improve the efficiency of resource scheduling and utilization. Summary of the Invention

[0005] The purpose of this invention is to provide a distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration, which solves the problems of task partitioning and resource scheduling being separated, insufficient modeling of task dependencies, and lack of real-time collaboration in hierarchical decision-making in the prior art, thereby reducing overall energy consumption and improving resource utilization efficiency and comprehensive performance.

[0006] To achieve the above objectives, this invention provides a distributed edge computing task scheduling method based on dynamic hypergraphs and hierarchical collaboration, comprising the following steps: Step S1: The gateway node obtains the task set, constructs a dynamic hypergraph model, and generates the system status by combining the system's real-time load indicators. Step S2: The gateway node generates task partitioning decisions based on the system state using a reinforcement learning model; Step S3: The scheduler node makes decisions based on task division and combines local observation information to make multi-agent collaborative decisions, and distributes the task subset to the server. Step S4: The server executes a task to calculate throughput and energy consumption in order to generate a latency feedback signal; Step S5: Based on the delayed feedback signal, optimize the reinforcement learning model and dynamically adjust the task partitioning decision.

[0007] Preferably, in step S1, constructing the dynamic hypergraph model specifically involves: Map tasks in the task set to vertices of a dynamic hypergraph; Construct a set of superedges; Assign attribute characteristics to the set of vertices and hyperedges.

[0008] Preferably, the constructed set of hyperedges includes: Based on the topological structure of the original task graph, vertices with direct predecessor or successor relationships are connected to construct order-dependent hyperedges. Connect all vertices with the same service type to construct a relational dependency hyperedge.

[0009] Preferably, attribute features are assigned to the vertex and hyperedge sets, specifically including: For each vertex, a comprehensive feature vector is constructed and attribute features are assigned, including the data size of the task, the type of service required, the latency tolerance, the relaxation time used to quantify the time urgency, the gateway transmission cost vector, and the dynamic computation cost vector. The attribute features assigned to sequentially dependent hyperedges include the communication cost matrix.

[0010] Preferably, in step S2, the reinforcement learning model employs a proximal policy optimization algorithm, which generates task partitioning decisions through an attention-based encoder-decoder network architecture. The encoder is constructed based on a hypergraph neural network, and the model is optimized through a composite reward function that includes structural optimization and cross-layer feedback.

[0011] Preferably, step S2 specifically includes: First, the system state of each frame is input into the encoder. The encoder uses a hypergraph neural network to extract high-order dependency features from the hypergraph structure information to obtain the hypergraph embedding. The hypergraph embedding is then concatenated and fused with real-time system metrics to generate a context vector. Based on the context vector, the decoder generates the partitioning action sequence of the current frame through an autoregressive approach, which determines which task to route to a specific scheduler, and finally generates the task partitioning decision.

[0012] Preferred composite reward function The specific expression is: ; in, Indicates a reward for communication energy efficiency; This indicates an encouragement to retain uncut hyperedges to save on cross-node communication costs; This indicates the energy consumption used to penalize transmission from the gateway to the scheduler; This indicates a load balancing penalty. This represents the load balancing metric calculated based on the standard deviation of the predicted load from each scheduler. This indicates a delayed feedback reward; This indicates the actual system throughput reported downstream at the end of the current frame.

[0013] Preferably, step S3 specifically includes: Step S301: Each scheduler node receives the assigned subset of tasks based on the task partitioning decision; Step S302: Each scheduler node acquires local observation information; Step S303: Each scheduler node makes multi-agent collaborative decisions based on local observation information and generates scheduling actions; Step S304: According to the scheduling action, schedule the task to the corresponding server for execution.

[0014] Preferably, all scheduler nodes are optimized by maximizing a shared team reward, specifically expressed as: ; in, This represents the total number of tasks newly completed by all schedulers within the current time slot; This represents the total energy consumption of the system due to calculations within the current time slot; and These represent the coefficients used to balance performance and energy consumption weights, respectively.

[0015] Preferably, the above method operates under a hierarchical collaborative workflow framework. The hierarchical collaborative workflow framework synchronizes request partitioning decision-making and request scheduling in a cyclical, multi-stage manner, and operates based on two time scales: a large time scale frame corresponding to the upper-layer gateway decision cycle, and a small time scale time slot corresponding to the lower-layer scheduler execution cycle.

[0016] Therefore, the distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration described above in this invention has the following beneficial technical effects: (1) This invention breaks through the limitation of traditional directed acyclic graphs in describing complex task dependencies. By introducing dynamic hypergraphs and associated dependency hyperedges, it can capture the implicit high-order affinity between tasks (such as shared service types or hardware preferences), enabling the gateway to identify and aggregate logically closely related tasks, avoiding the incorrect splitting and allocation of them to different scheduler nodes, thereby reducing resource fragmentation from the source and creating a more structured and predictable execution environment for the lower-level scheduler.

[0017] (2) This invention eliminates the problem of disconnect between upper-layer decisions and lower-layer resource status in traditional layered architectures. By constructing a cross-layer dual-loop feedback mechanism, the actual execution performance (such as throughput) of the lower-layer scheduler is converted into a delay feedback signal and fed back to the upstream gateway in real time. This mechanism breaks down the information barriers between the upper and lower layers, enabling the gateway to perceive the actual impact of its partitioning decisions on global performance and dynamically calibrate task partitioning decisions accordingly to adapt to real-time load changes of underlying resources. This effectively avoids policy mismatch, resource contention, and single-point overload caused by independent decision-making and lack of real-time perception.

[0018] (3) This invention realizes the transformation from single-stage optimization to system-level global optimization. Through the hierarchical collaborative workflow framework and composite reward function, the upper-level task partitioning decision and the lower-level scheduling decision are optimized simultaneously. This not only balances communication overhead and load balancing, but also integrates the dynamic hypergraph structure features with cross-layer delay feedback signals to construct a composite reward function that can guide the task partitioning decision learning toward long-term global optimum. This enables the entire scheduling process to reduce overall energy consumption while ensuring throughput and real-time requirements. Attached Figure Description

[0019] Figure 1 This is a flowchart of the distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration of the present invention; Figure 2 This is a diagram of the hierarchical collaboration workflow framework for the distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration of the present invention. Figure 3 The algorithm architecture for request layering; Figure 4 The experimental results are shown in the figure for comparison. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 This invention provides a distributed edge computing task scheduling method based on dynamic hypergraphs and hierarchical collaboration. This method operates within a hierarchical collaborative workflow framework. This framework synchronously performs request partitioning decisions and request scheduling in a cyclical, multi-stage manner, and operates based on two time scales: a large time-scale frame corresponding to the upper-layer gateway's decision cycle, and a small time-scale time slot corresponding to the lower-layer scheduler's execution cycle. Figure 2 As shown.

[0023] A frame consists of multiple consecutive time slots. At the beginning of each frame, the upper-layer gateway node performs a task partitioning decision, which remains valid throughout the entire frame period, providing stability for the lower-layer scheduling. The lower-layer scheduler node then performs distributed resource scheduling and execution based on the real-time status within each time slot that constitutes the frame.

[0024] like Figure 1 As shown, the distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration includes the following steps: Step S1: The gateway node obtains the task set, constructs a dynamic hypergraph model, and generates the system status by combining real-time system load metrics. Specifically: Map tasks in the task set to vertices of a dynamic hypergraph.

[0025] Constructing a set of hyperedges includes: Based on the topological structure of the original task graph, vertices with direct predecessor or successor relationships are connected to construct order-dependent hyperedges.

[0026] Connect all vertices with the same service type to construct a relational dependency hyperedge.

[0027] Assigning attribute characteristics to the set of vertices and hyperedges, including: For vertices Construct a comprehensive feature vector This vector integrates the static attributes and dynamic features of the task. The static attributes include the task's data size. Required service types Delay tolerance Dynamic features include relaxation time used to quantify time urgency. Gateway transmission cost vector and dynamically calculated cost vector .

[0028] Relaxation time The specific expression is: ; in, Index variable representing the task; Indicates the current task (i.e., vertex) The absolute deadline for the service request to be completed for the corresponding task. This indicates the real-time moment when the system constructs the current frame state.

[0029] Gateway transmission cost vector Used to predict the energy consumption of transmitting tasks from the gateway to each scheduler. Dynamically calculate the cost vector. Used to predict the execution energy consumption of this task under different schedulers. Dynamically calculate the cost vector. The j-th element The calculation formula is as follows: ; in, and These represent the average power coefficient and computing power of the cluster managed by the scheduler, respectively. Indicates the computational load of the task; This indicates the real-time load status of the scheduler; express of Power; A constant exponent representing the nonlinear relationship between computing power and power consumption; This represents the load impact factor, used to adjust the weighting of the scheduler's real-time load on the estimated execution energy consumption.

[0030] For sequentially dependent hyperedges The attribute characteristics assigned include a 3D communication cost matrix This aims to quantify the communication penalties incurred by cross-scheduler assignments. Represents the matrix dimension. Communication cost matrix. Each element in Defines when the parent task is assigned to the task with the number 1 The scheduler assigns subtasks to the numbered subtasks. The communication energy cost incurred by the scheduler is calculated using the following formula: ; in, Indicates the amount of output data between the parent and child tasks; and They represent the numbers respectively. The scheduler and its number are The transmission power and transmission rate between schedulers; This represents the penalty factor for big data transmission.

[0031] By integrating the aforementioned attribute features, the constructed hypergraph not only encodes the logical topology of the task, but also fully perceives the real-time heterogeneous state of the physical system.

[0032] In each frame At the initial moment, construct the system state. . It is constructed as a composite vector, designed to simultaneously capture macroscopic system load indicators and microscopic task structure features, with the specific expression as follows: ; in, Represents the set of all downstream schedulers; This refers to a specific scheduler within the set. ; This indicates the current real-time task queue length for all schedulers. The set is used to directly reflect the real-time congestion status of each edge cluster; This indicates the gateway buffer size, used to quantify the backlog of tasks waiting to be partitioned; This represents the set of hypergraph structure embeddings generated by the hypergraph neural network, which is the core part of the state and transforms non-Euclidean task dependencies into high-dimensional feature vectors.

[0033] Hypergraph Structure Embedded Set The generation process is achieved by a nonlinear mapping function. The description, specifically the expression, is: ; in, This represents the node feature matrix, which contains the attributes of all tasks to be processed; The hyperedge index tensor defines the topological connection structure of the hypergraph, that is, it specifies which task nodes are contained in which hyperedges; This represents the hyperedge attribute matrix, which quantifies the communication cost or association strength between task groups.

[0034] Step S2: The gateway node generates task partitioning decisions based on the system state using a reinforcement learning model.

[0035] The reinforcement learning model employs the Proximal Policy Optimization (PPO) algorithm, with the following architecture: Figure 3 As shown, a task partitioning decision is generated through an attention-based encoder-decoder network architecture. The encoder is built upon a hypergraph neural network, and the model is optimized using a composite reward function that incorporates structural optimization and cross-layer feedback. (Composite reward function) The specific expression is: ; in, Indicates a reward for communication energy efficiency; This indicates the savings in cross-node communication costs by encouraging the retention of uncut hyperedges (i.e., tasks within the same hyperedge are assigned to the same scheduler); This indicates the energy consumption used to penalize transmission from the gateway to the scheduler; This indicates a load balancing penalty. This represents a load balancing metric calculated based on the standard deviation of the predicted load of each scheduler, used to prevent single-point overload caused by excessive concentration of tasks. Delayed feedback rewards are key to achieving cross-layer collaboration; This represents the actual system throughput reported by the downstream system at the end of the current frame (i.e., the proportion of requests completed within the latency tolerance). It directly links the task partitioning decisions of the upper-layer gateway to the final global performance, forcing the task partitioning decisions to continuously adjust to adapt to the actual execution capabilities of the lower-layer regulator. By maximizing the various rewards of the cumulative composite reward function, joint optimization of energy consumption, load, and throughput is achieved. This represents a hyperedge in the hypergraph (i.e., an element in the set of hyperedges). Indicates the current frame The set of all "uncut" superedges in the array; Indicates a service request The first in One task; Indicates the current frame The set of all tasks to be partitioned within the scope.

[0036] like Figure 3 As shown, at the beginning of each frame, the system state of each frame is... Input to encoder. System status. Includes the queue length of each scheduler's real-time tasks. and gateway buffer size It also integrates a set of high-order hypergraph structure embeddings extracted from dynamic hypergraph topologies by hypergraph neural networks. (i.e., the node feature matrix, hyperedge index, and hyperedge attributes of a dynamic hypergraph).

[0037] The encoder uses a hypergraph neural network to extract high-order dependency features from the hypergraph structure information to obtain a hypergraph embedding. This hypergraph embedding is then concatenated and fused with real-time system metrics to generate a context vector. Based on this context vector, the decoder generates the segmentation action sequence for the current frame using an autoregressive approach. That is, deciding on each task Routing to the specific scheduler d (where ),like Figure 3 As shown, in the first decoding time step, the action is divided. This indicates that task number 5 will be... Assigned to scheduler number 3 In the second decoding time step, the action is divided. This indicates that task number 3 will be... Assigned to scheduler number 6 And so on, in the 1st Each decoding time step is used to divide the action. The task selected by the current action Assign to the scheduler selected in the current step. Finally, by linking all the partitioning actions together, a result based on the current system state is generated. Complete task partitioning strategy .

[0038] Step S3: The scheduler node makes decisions based on task partitioning and combines local observation information to perform multi-agent collaborative decision-making, distributing and scheduling task subsets to the server. Specifically, this includes: Step S301: Each scheduler node receives the assigned subset of tasks based on the task partitioning decision.

[0039] Step S302: Each scheduler node acquires local observation information. .in, This indicates the attributes of the currently scheduled task, including task data size, remaining deadline, service type, and dependency count; It indicates the real-time status of the server cluster managed by the scheduler node, including server availability and current computing power.

[0040] Step S303: Each scheduler node, based on local observation information, makes multi-agent collaborative decisions and generates scheduling actions. .

[0041] Local observation information The input is fed into a deep neural network policy model based on a multi-head attention mechanism for processing. Since the number of tasks in the waiting queue is dynamically changing, traditional fixed-input-dimension networks cannot handle it; therefore, a multi-head attention mechanism is employed. and Encoding is performed. The deep neural network policy model calculates the correlation strength between task features and server features, and finally outputs a probability distribution policy. .in, This represents the policy function, which is the mapping relationship learned by the deep neural network policy model; Indicates the current observation state Under these circumstances, specific scheduling actions will be taken. The conditional probability, which is a probability distribution strategy that covers the probability of selecting each feasible option in the action space.

[0042] Based on probability distribution strategy Generate the final scheduling action. .in, Scheduler The generated scheduling action is mathematically defined as a discrete index value that uniquely points to a valid "task-server pairing". Its decision logic is: if the action... Pointing to index , and index The corresponding meaning is "to move the tasks in the queue". Assigned to server If the scheduler executes the scheduling instruction immediately, the scheduler will do so.

[0043] Regarding the generation method, during the training phase, the scheduling action... Typically based on probability distribution strategies Random sampling is performed to explore different strategies; during the inference or actual execution phase, the index with the highest probability value is typically selected as the index. ,Right now To execute what is currently considered the optimal scheduling scheme. The variable representing the maximum value.

[0044] All scheduler nodes maximize a shared team reward. The optimization is performed, and the specific expression is as follows: ; in, This represents the total number of tasks newly completed by all schedulers within the current time slot; This represents the total energy consumption of the system due to calculations within the current time slot; and These represent the coefficients used to balance performance and energy consumption weights, respectively.

[0045] By maximizing this team reward, the distributed scheduler is incentivized to collaborate to complete more tasks and reduce energy consumption.

[0046] Step S304: According to the scheduling action, schedule the task to the corresponding server for execution.

[0047] Step S4: The server executes the task, monitors the task completion status and energy consumption data in real time, and calculates throughput and energy consumption to generate a latency feedback signal. Specifically, this includes: Each server executes its computational tasks according to the received scheduling instructions. The system records the actual completion time of each service request in real time. And the actual energy consumption during the execution process.

[0048] At the end of the current frame, statistics are calculated based on latency tolerance. The number of successfully completed service requests is used to calculate the system's real-time throughput, which serves as a latency feedback signal to the upper-layer gateway. The specific expression is as follows: ; in, Indicates the current frame The system's real-time throughput is used to measure the success rate of request processing within the current decision-making cycle, under the premise of meeting latency constraints. Indicates the current frame The set of service requests that arrive within the system and are processed by the system; The indicator function is a binary function that, when it satisfies The value is 1 if the condition is met, and 0 otherwise. Indicates a request The actual completion time is defined as the maximum value of the completion times of all subtasks in the DAG task graph of the request; Indicates the current frame The total number of requests reaching the gateway.

[0049] Step S5: Based on the delayed feedback signal, optimize the reinforcement learning model and dynamically adjust the task partitioning decision. Specifically, this includes: The system state of the current frame Divide the action sequence and the calculated composite reward function The data is stored in the experience replay buffer; the gradient is calculated using the PPO algorithm and the parameters of the encoder-decoder network are updated, so that the gateway tends to generate partitioning strategies that bring higher throughput and lower energy consumption in subsequent decisions, thus completing the reinforcement learning model update and dynamically adjusting the task partitioning decision for the next frame.

[0050] like Figure 3 As shown, this step employs an "actor-critic" dual-network architecture for collaborative training. The policy network acts as the "actor," corresponding to the encoder-decoder structure in the previous steps; its function is to receive the system state of the current frame. Output task partitioning strategy This generates specific actions, i.e., a task partitioning scheme for the current frame. Meanwhile, the value network acts as a "commentator," typically composed of multi-layered sensing mechanisms, whose function is to receive the same system state. Output a scalar value It is used to evaluate the expected cumulative reward of the current state (i.e., to assess the merits of the current situation), thereby calculating the advantage function to guide the optimization direction of the policy network.

[0051] To ensure the convergence stability of reinforcement learning during model parameter updates, a pruning mechanism unique to the PPO algorithm is introduced. Specifically, the system combines a composite reward function... When calculating the gradient, a pruning ratio is introduced when updating the policy network parameters. (In this embodiment, the value is 0.2), strictly limiting the probability ratio of the new and old strategies to within a certain range. Within the range. The physical significance of the pruning operation is to prevent the policy from oscillating violently due to excessively large single parameter updates, ensuring the monotonic non-decreasing of the policy iteration process (i.e., proximal optimization), thereby realizing a complete closed-loop learning process from actor-generated actions, environmental feedback rewards, to critic evaluation of value, and finally robust network updates through the pruning mechanism.

[0052] like Figure 4 As shown, to verify the effectiveness of the method of the present invention, comparative experiments were conducted with the present invention, a scheduling scheme based on a common directed acyclic graph, and a scheduling scheme without cross-layer feedback mechanism. Figure 4 In this diagram, Episode represents the training round (or iteration number) of reinforcement learning; Reward represents the cumulative reward value obtained by the agent in the current round, with a higher value indicating better performance. CHASE represents the method of this invention; W / o Hypergraph, or Without Hypergraph, represents a scheduling scheme based on a common directed acyclic graph, which removes the hypergraph structure from the state representation and uses only a common DAG graph and basic features; W / o Feedback, or Without Feedback, represents a scheduling scheme without cross-layer feedback mechanisms, which removes the delayed feedback signal from the lower-layer scheduler to the upper-layer gateway during training. Comparative experimental results show that the method of this invention has significant technical advantages in the following two aspects: (1) Compared with scheduling schemes based on ordinary directed acyclic graphs, the method of this invention breaks through the limitations of traditional state representation and improves the overall collaboration efficiency of the system. Traditional DAG models can only describe simple sequential dependencies between tasks, while the dynamic hypergraph introduced by the method of this invention can capture the implicit association attributes between tasks, such as whether they share the same service type or hardware resources, thereby generating task partitioning decisions with a stronger awareness of energy consumption. This high-order state representation not only aggregates logically closely related tasks from the source, effectively reducing cross-node communication overhead and resource waste, but also creates a more structured and predictable execution environment for the lower-level scheduler, improving the overall efficiency of hierarchical collaboration.

[0053] (2) Compared with scheduling schemes without cross-layer feedback mechanisms, the method of this invention eliminates the short-sightedness of upper-layer decisions and solves the execution bottleneck problem caused by resource status mismatch by introducing a cross-layer delayed feedback mechanism. Traditional schemes without feedback are prone to making short-sighted task allocation decisions due to their inability to perceive the real-time status of lower-layer resources, leading to overload of lower-layer nodes. The method of this invention transforms the actual execution effect of the lower-layer scheduler into a delayed feedback signal, which is fed back to the upper-layer gateway in real time. This prompts the gateway to learn the globally optimal strategy to adapt to the real underlying load, effectively breaking down the information barriers between the upper and lower layers and avoiding performance bottlenecks caused by the disconnect between decision-making and resource status. Thus, while ensuring throughput, the overall energy efficiency is maximized.

[0054] Therefore, the present invention adopts the above-mentioned distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration, which solves the problems of task partitioning and resource scheduling being separated, insufficient modeling of task dependency relationships, and lack of real-time collaboration in hierarchical decision-making in the prior art, thereby reducing overall energy consumption and improving resource utilization efficiency and comprehensive performance.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration, characterized in that, Includes the following steps: Step S1: The gateway node obtains the task set, constructs a dynamic hypergraph model, and generates the system status by combining the system's real-time load indicators. Step S2: The gateway node generates task partitioning decisions based on the system state using a reinforcement learning model; Step S3: The scheduler node makes decisions based on task division and combines local observation information to make multi-agent collaborative decisions, and distributes the task subset to the server. Step S4: The server executes a task to calculate throughput and energy consumption in order to generate a latency feedback signal; Step S5: Based on the delayed feedback signal, optimize the reinforcement learning model and dynamically adjust the task partitioning decision.

2. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 1, characterized in that, In step S1, the construction of the dynamic hypergraph model specifically involves: Map tasks in the task set to vertices of a dynamic hypergraph; Construct a set of superedges; Assign attribute characteristics to the set of vertices and hyperedges.

3. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 2, characterized in that, The constructed set of hyperedges includes: Based on the topological structure of the original task graph, vertices with direct predecessor or successor relationships are connected to construct order-dependent hyperedges. Connect all vertices with the same service type to construct a relational dependency hyperedge.

4. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 2, characterized in that, Assigning attribute characteristics to the set of vertices and hyperedges, specifically including: For each vertex, a comprehensive feature vector is constructed and attribute features are assigned, including the data size of the task, the type of service required, the latency tolerance, the relaxation time used to quantify the time urgency, the gateway transmission cost vector, and the dynamic computation cost vector. The attribute features assigned to sequentially dependent hyperedges include the communication cost matrix.

5. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 1, characterized in that, In step S2, the reinforcement learning model adopts a proximal policy optimization algorithm and generates task partitioning decisions through an attention-based encoder-decoder network architecture. The encoder is built based on a hypergraph neural network, and the model is optimized through a composite reward function that includes structural optimization and cross-layer feedback.

6. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 5, characterized in that, Step S2 specifically includes: First, the system state of each frame is input into the encoder. The encoder uses a hypergraph neural network to extract high-order dependency features from the hypergraph structure information to obtain the hypergraph embedding. The hypergraph embedding is then concatenated and fused with real-time system metrics to generate a context vector. Based on the context vector, the decoder generates the partitioning action sequence of the current frame through an autoregressive approach, which determines which task to route to a specific scheduler, and finally generates the task partitioning decision.

7. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 5, characterized in that, Composite reward function The specific expression is: ； in, Indicates a reward for communication energy efficiency; This indicates an encouragement to retain uncut hyperedges to save on cross-node communication costs; This indicates the energy consumption used to penalize transmission from the gateway to the scheduler; This indicates a load balancing penalty. This represents the load balancing metric calculated based on the standard deviation of the predicted load from each scheduler. This indicates a delayed feedback reward; This indicates the actual system throughput reported downstream at the end of the current frame.

8. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 1, characterized in that, Step S3 specifically includes: Step S301: Each scheduler node receives the assigned subset of tasks based on the task partitioning decision; Step S302: Each scheduler node acquires local observation information; Step S303: Each scheduler node makes multi-agent collaborative decisions based on local observation information and generates scheduling actions; Step S304: According to the scheduling action, schedule the task to the corresponding server for execution.

9. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 8, characterized in that, All scheduler nodes are optimized by maximizing a shared team reward, specifically expressed as: ； in, This represents the total number of tasks newly completed by all schedulers within the current time slot; This represents the total energy consumption of the system due to calculations within the current time slot; and These represent the coefficients used to balance performance and energy consumption weights, respectively.

10. The distributed edge computing task scheduling method based on dynamic hypergraph and hierarchical collaboration according to claim 1, characterized in that, The method operates within a hierarchical collaborative workflow framework, which synchronizes request partitioning decisions and request scheduling in a cyclical, multi-stage manner. It operates based on two time scales: a large time scale frame corresponding to the upper-layer gateway's decision cycle and a small time scale slot corresponding to the lower-layer scheduler's execution cycle.

Images