Multi-hop network topology-aware edge computing dependent task scheduling method and system

Abstract

The present application belongs to the field of edge computing and network technology, and specifically relates to a multi-hop network topology-aware dependent task scheduling method and system. In view of the problem that the existing scheduling ignores the multi-hop forwarding overhead of physical topology, the present scheme constructs a task logic directed acyclic graph and an edge physical network undirected graph, extracts features in parallel through an enhanced graph encoder, and performs cross-domain nonlinear fusion; a dynamic decision mechanism based on proximal policy optimization is designed, a virtual scheduling constraint edge is introduced to intelligently reconstruct the parallel relationship of tasks, and a dynamic mask is combined to ensure the correctness of causality; a multi-hop sensing heuristic scheduling is integrated at the lower level, and the optimized mapping of tasks to physical nodes is realized according to the shortest path and incremental cost evaluation. The present application effectively reduces the multi-hop transmission delay and system energy consumption, and significantly improves the overall execution efficiency of dependent tasks in the edge environment.

Description

Technical Field

[0001] This application relates to the field of edge computing and intelligent resource scheduling technology, and in particular to a multi-hop network topology-aware edge computing-dependent task scheduling method. Background Technology

[0002] With the rapid development of the Internet of Things (IoT), the Industrial Internet, and mobile smart terminals, massive amounts of data are continuously generated at the network edge, leading to the emergence of computationally intensive and latency-sensitive applications such as industrial video analytics, autonomous driving perception, and augmented reality. Traditional cloud computing relies on remote data centers to provide centralized computing power. While it has advantages in resource scale, it faces high communication latency and bandwidth overhead during wide area network transmission, making it difficult to meet the business requirements of low latency and high reliability. Edge computing, by pushing computing, storage, and service capabilities down to the network access side closer to the data source, shortens the service path at the system architecture level and has become a key paradigm supporting new industrial intelligent applications and the collaboration of massive numbers of terminals.

[0003] In practical industrial applications, complete applications typically consist of multiple subtasks with strict data dependencies and temporal constraints, and are generally modeled using directed acyclic graphs (DAGs). The introduction of dependencies transforms task scheduling from a simple resource mapping problem into a high-dimensional combinatorial optimization problem, requiring simultaneous satisfaction of topological causal constraints, heterogeneous computing power limitations of nodes, and multi-objective collaboration. Existing technologies for dependent task offloading mainly fall into two categories: traditional heuristic algorithms and intelligent scheduling methods based on deep reinforcement learning. However, these methods generally rely on the ideal assumption of fully interconnected edge servers and zero-latency cross-node data transmission when constructing system models and training frameworks.

[0004] This idealized assumption ignores the physical reality of servers being sparsely interconnected and having heterogeneous bandwidth in real-world wide-area or metropolitan edge deployments, typically via multi-hop links. When successive tasks in a directed acyclic graph are scheduled to different physical nodes, intermediate computation results must be forwarded via multi-hop routing through intermediate servers. Existing algorithms, lacking joint awareness of the underlying physical topology, often blindly assign data-intensive tasks to nodes with ample computing power but distant physical routes. While some studies have introduced multi-node collaboration, they have not performed deep state fusion between the task logical dependency graph and the edge physical network topology graph; the two-layer optimization architecture also lacks underlying multi-hop routing simulation feedback, resulting in large training reward variance and a sharp drop in the model's generalization ability in real sparse topologies.

[0005] In summary, existing technologies suffer from at least the following technical problems: First, they lack topology awareness. Current scheduling algorithms rely solely on task dependency graphs for state representation, failing to incorporate information such as connectivity and heterogeneity of link bandwidth in the edge physical network topology into the decision context. This leads to agents making erroneous decisions in sparse topologies, assigning tasks to computationally efficient but physically distant nodes. Second, multi-hop communication cost modeling is insufficient. The lower-layer heuristic scheduler does not dynamically trace the physical execution location of predecessor tasks and accumulate the shortest path transmission delay during node binding, resulting in unacceptable communication penalties in real multi-hop networks. Third, there is a disconnect between upper and lower layer optimization. The reward signal for reinforcement learning agents only reflects performance changes under ideal direct connection conditions, lacking realistic feedback from lower-level multi-hop routing simulations. This causes policy update directions to deviate from actual physical constraints, resulting in large reward variance and difficulty in convergence during training. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application aims to provide a multi-hop network topology-aware edge computing-dependent task scheduling method, which can effectively avoid the remote computing power trap, reduce cross-node communication overhead, and improve the global rationality and generalization ability of the scheduling scheme in non-fully interconnected multi-hop edge network environments.

[0007] To achieve the above objectives, this application adopts the following technical solution:

[0008] This application provides a multi-hop network topology-aware edge computing-dependent task scheduling method, the method comprising:

[0009] S1, Obtain the directed acyclic graph of the task logic; Obtain the undirected graph of the edge physical network; Initialize the policy network, value network, and augmented graph encoder parameters of the upper-layer near-end policy optimization;

[0010] S2 inputs the task logic directed acyclic graph and the edge physical network undirected graph in parallel into the augmented graph encoder; bidirectional graph convolution is used to aggregate messages along the forward and reverse edges of the task logic directed acyclic graph, and the concatenation is used to obtain the task node-level feature vector; a graph attention network is used to extract link congestion and node connectivity features of the edge physical network undirected graph; and a cross-domain attention pooling mechanism is used to nonlinearly fuse the computational load features of the task logic directed acyclic graph and the topological bottleneck features of the edge physical network undirected graph to output the global environment state feature vector.

[0011] S3, the actor network of the policy network receives the global environment state feature vector and outputs an action; the output action is to add a virtual scheduling constraint edge to the current task logic directed acyclic graph. This virtual scheduling constraint edge is used to force two subtasks that could originally be parallel to be transformed into a serial relationship with a sequential execution order. A dynamic action masking mechanism is used to ensure that the addition of the virtual scheduling constraint edge does not destroy the strict causal execution order of the original directed acyclic graph, and an enhanced logic graph is generated.

[0012] S4 receives the enhanced logic graph and the undirected graph of the edge physical network, performs multi-hop perceptual heuristic scheduling, and outputs a physical offloading scheme.

[0013] S5: Calculate the system multi-objective joint performance value including multi-hop forwarding costs based on the physical offloading scheme; define the immediate reward as the relative change between the system multi-objective joint performance value corresponding to the augmented logic graph and the system multi-objective joint performance value before adding virtual scheduling constraint edges; calculate the temporal difference residual and estimate the cumulative advantage function through generalized advantage; sample batch data from the experience replay pool, update the actor network using the probability proportional pruning objective function of near-end policy optimization, and simultaneously update the value network by minimizing the mean square error; repeat steps S2 to S5 until the policy network converges.

[0014] As a preferred technical solution, in step S1, the enhanced graph encoder adopts a dual-flow graph encoding structure:

[0015] The task graph encoding stream uses a bidirectional graph convolutional network, which includes a forward message passing layer and a backward message passing layer. The forward layer aggregates predecessor features along the direction from the start node to the end node, and the backward layer aggregates successor features along the direction from the end node to the start node. The bidirectional aggregation results are concatenated and then nonlinearly transformed to obtain task node-level features.

[0016] The physical topology graph encoding stream uses a multi-head graph attention network to calculate the attention coefficients of each edge node with all its neighboring nodes, weighted aggregate the link bandwidth and computing resource characteristics of the neighboring nodes, and obtain the physical node-level features after multi-head splicing and averaging.

[0017] In step S2, the cross-domain attention pooling uses task node-level features as query vectors and physical node-level features as key-value vectors to calculate cross-domain attention weights and perform weighted pooling on physical features. Then, it is concatenated with the global pooling features of the task graph and mapped through a fully connected layer to output a global environment state feature vector of fixed dimensions.

[0018] As a preferred technical solution,

[0019] In step S3, the dynamic action masking mechanism imposes three rigid constraints on the candidate virtual edges: the non-repetition constraint requires that the candidate edge has a value of zero in the current graph adjacency matrix; the non-redundancy constraint requires that the reachability of the candidate edge from the start node to the end node in the current graph transitive closure matrix is ​​zero; and the strict topological order constraint requires that the index of the start node in the current graph topological sort is less than the index of the end node.

[0020] The output action space of the actor network is the probability distribution on the set of legal candidate edges that satisfy all the above constraints. The actual virtual edges added are determined by sampling.

[0021] As a preferred technical solution,

[0022] In step S4, the enhanced logic graph and the undirected graph of the edge physical network are received, multi-hop awareness heuristic scheduling is performed, and the output physical offloading scheme includes:

[0023] Maintain a global ready task queue and push all task nodes with an in-degree of zero into the queue;

[0024] Calculate the underlying priority value of each task in the ready task queue and rearrange them in descending order;

[0025] Pop up the first task in the queue and filter the set of candidate physical nodes that meet the constraints of the operating environment;

[0026] For each candidate physical node, trace the set of physical execution locations of all direct predecessor tasks of this task, call the shortest path algorithm to query the shortest route path from the physical execution location of the predecessor task to the current candidate physical node in the undirected graph of the edge physical network, accumulate the transmission delay of each link on the shortest route path based on the link communication rate to calculate the multi-hop forwarding cost, and combine the local computing delay, queuing delay and energy consumption cost of the current candidate physical node to predict the incremental cost of the scheduling decision on the system's comprehensive objective function;

[0027] Select the physical node with the lowest incremental cost to bind the current task, update the idle timestamp of the physical node, push the subsequent tasks that meet the dependency conditions into the ready task queue, and execute in a loop until all tasks are scheduled and output the physical unloading solution.

[0028] As a preferred technical solution,

[0029] In step S4, the multi-hop forwarding cost is calculated as follows: when the bound physical node of the direct predecessor task of the task to be scheduled is the same as the current candidate node, the cost is zero; when they are different nodes, the shortest path algorithm is called to query the shortest route from the predecessor node to the candidate node in the physical graph. The multi-hop forwarding cost is equal to the sum of the transmission delays obtained by dividing the output data of all links on the path by the corresponding communication rate.

[0030] As a preferred technical solution,

[0031] In step S4, the incremental cost is a weighted sum of multi-hop forwarding cost, local computation latency, queuing latency, and energy consumption cost; local computation latency equals the task computation amount divided by the candidate node's computation capacity; queuing latency equals the positive part of the difference between the candidate node's current idle timestamp and the latest completion time of all the task's direct predecessor tasks; energy consumption cost equals the candidate node's unit energy consumption coefficient multiplied by the task computation amount; each item is multiplied by its corresponding weight coefficient and then summed to obtain the predicted incremental cost of the candidate node.

[0032] As a preferred technical solution,

[0033] In step S5, the system multi-objective joint performance value is a weighted combination of total processing latency, total system energy consumption, total execution cost, and application hit rate. The total processing latency is the sum of the computation latency of all tasks and the multi-hop forwarding latency of all cross-node dependent edges. The application hit rate is the proportion of the number of applications that successfully complete all sub-task scheduling within a preset deadline to the total number of applications. The immediate reward is the difference between the performance value corresponding to the enhanced graph and the performance value before adding the edge. The reward is positive when the performance value after enhancement is better, and negative otherwise.

[0034] As a preferred technical solution,

[0035] In step S5, the objective function is: calculate the probability ratio of the new strategy and the old strategy in the sampling action, trim the probability ratio to a preset interval, take the product of the trimmed ratio and the dominance function, and then take the minimum value of the product of the untrimmed ratio and the dominance function as the optimization objective; the value network is updated by minimizing the mean square error of the current value network output by minimizing the temporal difference objective value; the dominance estimation adopts generalized dominance estimation, which is a weighted sum of multiple temporal difference errors and the cumulative dominance function.

[0036] As a preferred technical solution,

[0037] The nodes of the directed acyclic graph of the task logic include computational and data volume attributes, task type labels, and deadline constraints.

[0038] The edge attributes of the undirected graph of the edge physical network include link communication rate, link latency, and reliability metrics.

[0039] The enhanced graph encoder parameters are initialized uniformly using Xavier;

[0040] The parameters of the actor network and value network are orthogonally initialized;

[0041] The experience replay pool adopts priority experience replay, and the sampling priority is allocated according to the size of the time difference error. The larger the error, the higher the probability of the sample being sampled.

[0042] This application also provides a multi-hop network topology-aware edge computing-dependent task scheduling system, the system comprising:

[0043] The dual-graph construction module is used to obtain the task logic directed acyclic graph and the edge physical network undirected graph, and to perform structured storage of the node attributes of the task logic directed acyclic graph and the node and edge attributes of the edge physical network undirected graph.

[0044] The state encoding module includes an enhanced graph encoder, which encodes the task logic directed acyclic graph and the edge physical network undirected graph in parallel. It extracts task graph features and physical topology graph features through bidirectional graph convolution and graph attention network, respectively, and fuses them through cross-domain attention pooling mechanism to output a global environment state feature vector.

[0045] The policy sampling module includes an actor network and a value network for the near-end policy optimization policy network. The actor network is used to receive the global environment state feature vector and output the actions with added virtual scheduling constraint edges. The value network is used to evaluate the value of the current state. The policy sampling module also includes a dynamic action mask unit, which is used to verify the legality of the actions output by the actor network.

[0046] The multi-hop scheduling module is used to receive the augmented logic graph and the undirected graph of the edge physical network, and perform multi-hop awareness heuristic scheduling. The multi-hop scheduling module includes a ready queue management unit, a low-level priority calculation unit, a predecessor position tracing unit, a shortest path query unit, a multi-hop forwarding cost calculation unit, and an incremental cost evaluation unit, which are used to bind each task in the augmented logic graph to the physical node that minimizes the incremental cost and output a physical offloading scheme.

[0047] The reward update module is used to calculate the joint performance value of the system with multi-hop forwarding costs based on the physical offloading scheme, calculate the immediate reward, estimate the cumulative advantage function through generalized advantage, and update the actor network and value network using the probabilistic pruning objective function optimized by the near-end policy.

[0048] Compared with the prior art, the beneficial effects of this application are as follows:

[0049] This application achieves joint state representation of the two graphs by parallel encoding of the directed acyclic graph of the task logic and the undirected graph of the edge physical network through an enhanced graph encoder and cross-domain attention fusion. This enables the reinforcement learning agent to not only perceive the computational requirements and dependency structure of the task when making decisions, but also to simultaneously identify the connectivity and potential routing bottlenecks of the underlying physical network. This eliminates the blind decision to allocate high-data-volume tasks to congested nodes at the state space level and improves the global rationality of the scheduling scheme.

[0050] This application introduces a predecessor location backtracking and shortest path query module into the lower-level heuristic scheduler, accurately accumulating the cross-node multi-hop transmission delay into the node binding cost prediction function. This achieves explicit modeling and white-box calculation of multi-hop routing costs, effectively avoiding the remote computing power trap and ensuring the global rationality of the scheduling scheme in sparse topologies.

[0051] This application constructs a complete data flow consisting of upper-layer topology exploration, lower-layer multi-hop routing simulation, and weighted performance reward feedback, enabling reinforcement learning agents to update policy gradients based on real communication penalties. This solves the problem of reward signal distortion, reduces training reward variance, and accelerates policy network convergence.

[0052] This application uses a topology-aware closed-loop training architecture to enable the model to learn the intrinsic mapping rules between topological features and scheduling costs rather than memorizing fixed-size data structures. It exhibits excellent zero-shot generalization ability in real industrial scenarios with cross-scale and cross-topology, and has engineering deployment adaptability. Attached Figure Description

[0053] Figure 1 A schematic diagram of the physical deployment structure for edge computing-dependent task scheduling methods that are topology-aware in multi-hop networks;

[0054] Figure 2 A schematic diagram of the topology-aware two-layer optimization framework for multi-hop route-aware lower-layer physical scheduling (TA-BiRLH);

[0055] Figure 3 A comparison chart of system target values ​​under different server sizes;

[0056] Figure 4 To provide a hit rate comparison chart;

[0057] Figure 5 A graph showing the trend of the target value as user load increases;

[0058] Figure 6 This is a chart comparing application hit rates.

[0059] Figure 7 A bar chart showing the target values ​​under actual load;

[0060] Figure 8 To apply a hit rate histogram;

[0061] Figure 9 A comparison chart of target values ​​across different scales;

[0062] Figure 10 This is a chart comparing hit rates across different scales. Detailed Implementation

[0063] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0064] This application provides a multi-hop network topology-aware edge computing dependent task scheduling method, which models the edge computing dependent task offloading problem as a topology-aware two-layer optimization architecture, and implements it through an iterative process of "joint state representation → upper-layer structure exploration → lower-layer multi-hop scheduling → closed-loop reward update".

[0065] like Figure 1 The diagram illustrates the applicable scenarios and physical deployment structures for this method. Figure 1 The lower half is the terminal device layer, where each device generates an application containing multiple subtasks. These subtasks are organized in a DAG (such as nodes) structure. (etc.), the arrows indicate the direction of data dependency. The upper part of the diagram represents the edge server layer, the server... and Solid lines indicate multi-hop physical communication links. Task unloading paths are indicated by dashed arrows: red dashed lines represent tasks. Unload to edge server Execution, the purple dashed line represents the task. Unload to edge server Execution. The current driver task is on the server. After execution, its output needs to be forwarded to the server where the subsequent task is located via multi-hop routing through intermediate nodes. This architecture forms a cross-node result return path. It realistically depicts the physical constraints of "terminal generating dependent tasks → edge node distributed execution → multi-hop link result return", which is the basic scenario for the explicit modeling of multi-hop communication costs in this invention.

[0066] like Figure 2 As shown, this application employs a multi-hop route-aware lower-level physical scheduling (TA-BiRLH) topology-aware two-layer optimization framework. To overcome the "topology blindness" defect of traditional scheduling algorithms in multi-hop environments, this invention constructs a two-layer optimization closed loop with upper and lower layer collaboration. The upper layer uses a reinforcement learning agent to structurally enhance the task logic graph, while the lower layer performs heuristic scheduling under real physical topology constraints and feeds back the global performance after multi-hop communication penalty calculation to the upper layer, driving the policy network convergence.

[0067] Figure 2 This paper demonstrates the data interaction and training closed loop of the two-layer optimization architecture of this invention. The framework consists of two layers: the upper layer is a topology-aware reinforcement learning policy, including an augmented state encoder and a PPO policy network. The encoder receives the current task logic graph (DAG). The system outputs a joint state representation based on the edge physical topology. The policy network then generates an augmented graph based on this state output, which represents the action of "adding a legal dependency edge" (indicated by the red arrow pointing from the brown node to the yellow node). The lower layer is a multi-hop awareness heuristic scheduler that receives the enhanced logical graph and physical topology, executes the TA-CPLS node binding algorithm, and outputs a physical offloading scheme. and The upper and lower levels are connected via a reward formula. A closed loop is formed: the lower layer calculates the comprehensive objective function value including multi-hop communication penalties, and feeds the difference as an immediate reward to the upper layer to guide the PPO network in policy gradient updates. This framework realizes a topology-aware training paradigm of "logical structure exploration → physical route simulation → performance reward feedback".

[0068] The multi-hop network topology-aware edge computing-dependent task scheduling method of this application includes the following steps:

[0069] S1, obtain the directed acyclic graph of the task logic; obtain the undirected graph of the edge physical network; initialize the policy network, value network and enhanced graph encoder (EGE) parameters of the upper-layer near-end policy optimization (PPO);

[0070] Among them, the task logic is a directed acyclic graph. Node attributes include computational cost. Input data volume Output data volume And service type tags.

[0071] Undirected graph of edge physical network Nodes represent sets of edge servers, and edges represent physical communication links and their corresponding communication rates. .

[0072] S2, the directed acyclic graph of task logic Undirected graphs of edge physical networks A parallel input augmented graph encoder is used; bidirectional graph convolution is applied to the directed acyclic graph of the task logic to aggregate messages along both forward and reverse edges, and the concatenation is used to obtain the task node-level feature vector; a graph attention network is used to extract link congestion and node connectivity features from the undirected graph of the edge physical network; and a cross-domain attention pooling mechanism is used to nonlinearly fuse the computational load features of the directed acyclic graph of the task logic with the topological bottleneck features of the undirected graph of the edge physical network, outputting a global environment state feature vector. ;

[0073] S3, the actor network of the policy network, receives the global environment state feature vector and outputs actions. This output action To add a virtual scheduling constraint edge to the current task logic directed acyclic graph, this virtual scheduling constraint edge is used to force two originally parallel subtasks into a serial relationship with a sequential execution order; a dynamic action masking mechanism is used to ensure that the addition of the virtual scheduling constraint edge does not disrupt the strict causal execution order of the original directed acyclic graph, thus generating an enhanced logic graph.

[0074] It's important to note that the virtual scheduling constraint edge is not an application of existing business data dependencies, but rather a kind of affinity binding constraint. Its physical meaning lies in this: when the reinforcement learning agent anticipates that scheduling two subtasks that could originally be parallel to nodes that are physically far apart or have limited bandwidth would incur high multi-hop communication penalties, the agent proactively adds this virtual edge, forcing the two tasks to logically transform into a sequential relationship with a predetermined execution order. This effectively reduces the ineffective scheduling search space of the underlying heuristic algorithm and avoids the "far-end computing power trap."

[0075] S4 receives the enhanced logic graph and the undirected graph of the edge physical network, performs multi-hop perceptual heuristic scheduling, and outputs a physical offloading scheme.

[0076] S5, Calculate the joint multi-objective performance value of the system including multi-hop forwarding costs based on the physical offloading scheme. The immediate reward is defined as the relative change in the system's multi-objective joint performance value corresponding to the enhanced logic graph compared to the system's multi-objective joint performance value before adding virtual scheduling constraint edges. ; Calculate the temporal difference (TD) residuals and accumulate the advantage function through generalized advantage estimation (GAE); sample batch data from the experience replay pool, update the actor network using the probabilistic proportional pruning objective function of proximal policy optimization (PPO), and simultaneously update the value (also known as: critic) network by minimizing the mean square error; repeat steps S2 to S5 until the policy network converges.

[0077] As a preferred technical solution, in step S1, the enhanced graph encoder adopts a dual-flow graph encoding structure:

[0078] The task graph encoding stream uses a bidirectional graph convolutional network, which includes a forward message passing layer and a backward message passing layer. The forward layer aggregates predecessor features along the direction from the start node to the end node, and the backward layer aggregates successor features along the direction from the end node to the start node. The bidirectional aggregation results are concatenated and then nonlinearly transformed to obtain task node-level features.

[0079] The physical topology graph encoding stream uses a multi-head graph attention network to calculate the attention coefficients of each edge node with all its neighboring nodes, weighted aggregate the link bandwidth and computing resource characteristics of the neighboring nodes, and obtain the physical node-level features after multi-head splicing and averaging.

[0080] In step S2, the cross-domain attention pooling uses task node-level features as query vectors and physical node-level features as key-value vectors to calculate cross-domain attention weights and perform weighted pooling on physical features. Then, it is concatenated with the global pooling features of the task graph and mapped through a fully connected layer to output a fixed-dimensional global environment state feature vector, which serves as the decision context for the upper-layer policy network.

[0081] As a preferred technical solution,

[0082] In step S3, to ensure that adding virtual edges does not disrupt the strict causal execution order of the original task logic directed acyclic graph, a dynamic action masking mechanism is introduced into the action space to forcibly filter out illegal actions.

[0083] The dynamic action masking mechanism for candidate virtual edges Apply three rigid constraints:

[0084] The non-repetition constraint requires that candidate edges have a value of zero in the current graph's adjacency matrix;

[0085] The non-redundancy constraint requires that the reachability of a candidate edge from its starting node to its ending node in the current graph's transitive closure matrix be zero (i.e., there are no other indirectly connected paths).

[0086] Strict topological ordering constraints require that the index of the starting node in the current graph topological sort is less than the index of the ending node;

[0087] The algorithm only allows pointers from nodes with a "topologically earlier" order to nodes with a "topologically later" order during decision-making. This completely eliminates the possibility of cycles from a graph theory perspective, ensuring the enhanced graph... It remains a valid directed acyclic graph.

[0088] The output action space of the actor network is the probability distribution on the set of legal candidate edges that satisfy all the above constraints. The actual virtual edges added are determined by sampling.

[0089] As a preferred technical solution,

[0090] In step S4, the enhanced logic graph and the undirected graph of the edge physical network are received, multi-hop awareness heuristic scheduling is performed, and the output physical offloading scheme includes:

[0091] Maintain a global ready task queue and push all task nodes with an in-degree of zero into the queue;

[0092] Calculate the underlying priority value of each task in the ready task queue and rearrange them in descending order;

[0093] Pop up the first task in the queue and filter the set of candidate physical nodes that meet the constraints of the operating environment;

[0094] For each candidate physical node, trace the set of physical execution locations of all direct predecessor tasks of this task, call the shortest path algorithm to query the shortest route path from the physical execution location of the predecessor task to the current candidate physical node in the undirected graph of the edge physical network, accumulate the transmission delay of each link on the shortest route path based on the link communication rate to calculate the multi-hop forwarding cost, and combine the local computing delay, queuing delay and energy consumption cost of the current candidate physical node to predict the incremental cost of the scheduling decision on the system's comprehensive objective function;

[0095] Select the physical node with the lowest incremental cost to bind the current task, update the idle timestamp of the physical node, push the subsequent tasks that meet the dependency conditions into the ready task queue, and execute in a loop until all tasks are scheduled and output the physical unloading solution.

[0096] As a preferred technical solution,

[0097] In step S4, the multi-hop forwarding cost is calculated as follows: when the bound physical node of the direct predecessor task of the task to be scheduled is the same as the current candidate node, the cost is zero; when they are different nodes, the shortest path algorithm is called to query the shortest route from the predecessor node to the candidate node in the physical graph. The multi-hop forwarding cost is equal to the sum of the transmission delays obtained by dividing the output data of all links on the path by the corresponding communication rate.

[0098] As a preferred technical solution,

[0099] In step S4, the incremental cost is a weighted sum of multi-hop forwarding cost, local computation latency, queuing latency, and energy consumption cost; local computation latency equals the task computation amount divided by the candidate node's computation capacity; queuing latency equals the positive part of the difference between the candidate node's current idle timestamp and the latest completion time of all the task's direct predecessor tasks; energy consumption cost equals the candidate node's unit energy consumption coefficient multiplied by the task computation amount; each item is multiplied by its corresponding weight coefficient and then summed to obtain the predicted incremental cost of the candidate node.

[0100] Specifically, a global ready task queue (containing all tasks with an in-degree of 0) is maintained using multi-hop route-aware lower-level physical scheduling (TA-CPLS). The bottom-level priority value (Bottom-Level Rank) of each task in the queue is calculated and rearranged in descending order. The task at the head of the queue is popped, and a set of candidate physical nodes that meet the runtime environment constraints is selected. For each candidate node implement:

[0101] 1. Tracing the predecessor location: Tracing the set of physical execution locations of all direct predecessor tasks of this task. ;

[0102] 2. Query the shortest path: Call the shortest path algorithm to query the edge physical topology routing table and obtain the shortest path. to Shortest route path ;

[0103] 3. Accumulated multi-hop latency: based on the formula Accumulate the transmission delay of each link on the path and calculate the cost of multi-hop forwarding across nodes;

[0104] 4. Predict incremental cost: Combining the target node's computational latency, queuing time, and energy consumption cost, predict the incremental cost of this scheduling decision on the system's overall objective function. Select the physical node with the minimum incremental cost to bind the task, update the node's idle timestamp, unlock subsequent tasks that meet the dependency conditions and enqueue them, and repeat this process until all tasks in the system are cleared.

[0105] As a preferred technical solution,

[0106] In step S5, the system multi-objective joint performance value is a weighted combination of total processing latency, total system energy consumption, total execution cost, and application hit rate. The total processing latency is the sum of the computation latency of all tasks and the multi-hop forwarding latency of all cross-node dependent edges. The application hit rate is the proportion of the number of applications that successfully complete all sub-task scheduling within a preset deadline to the total number of applications. The immediate reward is the difference between the performance value corresponding to the enhanced graph and the performance value before adding the edge. The reward is positive when the performance value after enhancement is better, and negative otherwise.

[0107] Specifically, the following discloses the specific algorithm for multi-hop route-aware lower-level physical scheduling (TA-CPLS):

[0108] Among them, the multi-hop route-aware critical path local selection scheduling algorithm (TA-CPLS) inputs: augmented task DAG Edge physical topology map Link rate matrix Task attribute set Output: Task physical unloading decision scheme System overall performance value .

[0109] 1: Initialize the ready task queue (ready_list) and set it to... Push all task nodes with an in-degree of 0 into the queue.

[0110] 2: while ready_list do

[0111] 3: Calculate the bottom-level absolute priority value of each task in ready_list and rearrange them in descending order.

[0112] 4: Pop up the highest priority task at the top of the queue.

[0113] 5: Based on the edge node environment configuration, construct a set of candidate physical nodes that meet the operational constraints.

[0114] 6: for each candidate node do

[0115] 7: Initialize the predicted incremental cost for this node.

[0116] 8: / / Explicit modeling of core multi-hop cost

[0117] 9: for task Each direct precursor mission do

[0118] 10: Trace the physical execution location of the precursor mission. 11: if then

[0119] 12: Call the routing algorithm to query middle to Shortest route path

[0120] 13: Calculate the multi-hop forwarding latency penalty:

[0121] 14:

[0122] 15: end if

[0123] 16: end for

[0124] 17: Compute Node The local computing latency, queuing latency, and energy consumption costs are added together to

[0125] 18: end for

[0126] 19: Choose to make smallest physical node ,Will Bind to

[0127] 20: Update node The idle timestamp, if If all predecessors of a task have been completed, then push it into the ready_list.

[0128] 21: end while

[0129] 22: Based on the final scheduling scheme Calculate the joint multi-objective performance of the system including multi-hop communication penalties. And return.

[0130] As a preferred technical solution,

[0131] In step S5, the objective function is: calculate the probability ratio of the new strategy and the old strategy in the sampling action, trim the probability ratio to a preset interval, take the product of the trimmed ratio and the dominance function, and then take the minimum value of the product of the untrimmed ratio and the dominance function as the optimization objective; the value network is updated by minimizing the mean square error of the current value network output by minimizing the temporal difference objective value; the dominance estimation adopts generalized dominance estimation, which is a weighted sum of multiple temporal difference errors and the cumulative dominance function.

[0132] As a preferred technical solution,

[0133] The nodes of the directed acyclic graph of the task logic include computational and data volume attributes, task type labels, and deadline constraints.

[0134] The edge attributes of the undirected graph of the edge physical network include link communication rate, link latency, and reliability metrics.

[0135] The enhanced graph encoder parameters are initialized uniformly using Xavier;

[0136] The parameters of the actor network and value network are orthogonally initialized;

[0137] The experience replay pool adopts priority experience replay, and the sampling priority is allocated according to the size of the time difference error. The larger the error, the higher the probability of the sample being sampled.

[0138] It should be noted that the bidirectional GCN and cross-domain attention pooling in step S2 can be replaced by graph attention network (GAT), GraphSAGE or graph Transformer structure, which can also achieve node neighborhood feature aggregation and joint representation of the two graphs.

[0139] It should be noted that the shortest path algorithm in step S4 can be replaced by the A* search algorithm, restricted Dijkstra's algorithm, or a pre-computed routing table caching mechanism in large-scale topologies to balance online scheduling latency and path accuracy.

[0140] It should be noted that the PPO algorithm in steps S3 and S5 can be replaced by soft Actor-Critic (SAC), deep deterministic policy gradient (DDPG), or the distributed IMPALA framework, which can also achieve policy optimization in continuous / discrete action spaces.

[0141] It should be noted that the underlying critical path sorting of the TA-CPLS scheduling logic in step S4 can be replaced by the HEFT (Heterogeneous Earliest Finish Time) algorithm or Dynamic Priority Scheduling (DPS). As long as the multi-hop cost accumulation module is integrated, the same technical effect can be achieved.

[0142] It should be noted that although this invention models edge topology using undirected graphs, it can be equivalently extended to directed graphs, time-varying graphs, or software-defined networking (SDN) control plane topologies. As long as link status queries and shortest path calculations are supported, the core scheduling logic remains unchanged.

[0143] This application also provides a multi-hop network topology-aware edge computing-dependent task scheduling system, the system comprising:

[0144] The dual-graph construction module is used to obtain the task logic directed acyclic graph and the edge physical network undirected graph, and to perform structured storage of the node attributes of the task logic directed acyclic graph and the node and edge attributes of the edge physical network undirected graph.

[0145] The state encoding module includes an enhanced graph encoder, which encodes the task logic directed acyclic graph and the edge physical network undirected graph in parallel. It extracts task graph features and physical topology graph features through bidirectional graph convolution and graph attention network, respectively, and fuses them through cross-domain attention pooling mechanism to output a global environment state feature vector.

[0146] The policy sampling module includes an actor network and a value network for the near-end policy optimization policy network. The actor network is used to receive the global environment state feature vector and output the actions with added virtual scheduling constraint edges. The value network is used to evaluate the value of the current state. The policy sampling module also includes a dynamic action mask unit, which is used to verify the legality of the actions output by the actor network.

[0147] The multi-hop scheduling module is used to receive the augmented logic graph and the undirected graph of the edge physical network, and perform multi-hop awareness heuristic scheduling. The multi-hop scheduling module includes a ready queue management unit, a low-level priority calculation unit, a predecessor position tracing unit, a shortest path query unit, a multi-hop forwarding cost calculation unit, and an incremental cost evaluation unit, which are used to bind each task in the augmented logic graph to the physical node that minimizes the incremental cost and output a physical offloading scheme.

[0148] The reward update module is used to calculate the joint performance value of the system with multi-hop forwarding costs based on the physical offloading scheme, calculate the immediate reward, estimate the cumulative advantage function through generalized advantage, and update the actor network and value network using the probabilistic pruning objective function optimized by the near-end policy.

[0149] To further clarify the technical details and execution process of the present invention, a specific numerical simulation embodiment is provided below to demonstrate the system completing one complete iteration (S1). The operating mechanism of S5.

[0150] 1. Scene initialization (corresponding to S1)

[0151] Edge physical topology Includes 4 edge servers The nodes are linearly connected: 1-2-3-4. The link communication rate between adjacent nodes is [missing information]. .

[0152] Task Logic DAG It contains 6 sub-tasks. The initial dependencies are: , , , , , .

[0153] Task attributes: To simplify calculations, the computational cost of all tasks is assumed to be... cycles, output data volume All edge servers have a computing power of [missing information]. (That is, the computation time for a single task is always the same) ).

[0154] 2. Dual-graph state representation (corresponding to S2)

[0155] Enhanced graph encoder (EGE) receives the above and Through graph convolution aggregation, the state vector Key information captured: mission These are parallel branches, and if they are assigned to physically distant nodes (such as node 1 and node 4), they converge to... This will result in high multi-hop transmission costs.

[0156] 3. Optimization of upper-level topology (corresponding to S3)

[0157] PPO policy network is state-based Sampling action. Agent decision-making requires constraints. and To reduce potential communication risks, the parallelism is increased, and the output action is to add virtual scheduling constraint edges. Acyclicity check: In the current topological sort, The level is lower than Furthermore, there is no existing path between the two edges. This edge satisfies the three constraints of non-repetition, non-redundancy, and topological order, and is therefore legally allowed. Generate an enhanced graph. ,at this time Must You can only begin after this is completed.

[0158] 4. Lower-layer multi-hop awareness scheduling (corresponding to S4, TA-CPLS executes Trace)

[0159] Scheduler based on Execute Algorithm 1:

[0160] Scheduling The ready queue contains only Candidate nodes With no predecessor, the cost of multiple hops is 0. Greedily select node 1. Completion Time Record location .

[0161] Scheduling The ready queue contains .

[0162] If node 1 is selected: Pre-driver At node 1, the cost of cross-node prediction within the same node is 0. Total prediction cost = Calculation .

[0163] If node 4 is selected: Pre-driver At node 1, query the shortest path. Multi-hop cost Total forecast cost = .

[0164] Decision: Select node 1, which has the lowest cost. , .

[0165] Scheduling The ready queue contains (Due to virtual edges) We need to wait Finish).

[0166] If node 1 is selected: Pre-driver At node 1, the cost across nodes is 0. Start time Completion Time .

[0167] If node 2 is selected: Precursor At node 1. Path Multi-hop cost Start time Completion Time .

[0168] Decision: Select node 1, which has the lowest cost. , .

[0169] Scheduling The ready queue contains Suppose that at this point, node 1 experiences a surge in queuing latency due to the intervention of other high-priority tasks (e.g., queuing time increases significantly). ), while node 3 is idle.

[0170] If node 3 is selected: Pre-drive All are on node 1. Query the shortest path. Multi-hop cost .

[0171] Readiness time = .

[0172] Completion Time .

[0173] Decision: After comprehensive consideration, the total incremental cost of node 3 is still better than that of the congested node 1, therefore, node 3 is bound. (To be continued) Continue in this manner until the queue is empty, then output the final solution. )

[0174] 5. Closed-loop reward feedback (corresponding to S5) lower-level scheduler output scheme Then, calculate the overall system performance value including the aforementioned explicit multi-hop penalty. Assuming the baseline scheme before adding edges will... Distributed scheduling to nodes 1 and 4 resulted in... The aggregation process generates a huge communication delay, and its efficiency value is The current enhancement scheme has an effectiveness value of [value missing]. (A smaller value indicates a lower overall cost and better performance). Calculate immediate reward: This positive reward Feedback is sent to the upper-layer PPO network, and combined with GAE advantage estimation and Clipped objective function, the policy network weights are updated, prompting the agent to be more inclined to output such effective "virtual scheduling constraint edges" when encountering similar topological bottlenecks in the future.

[0175] The present invention has the following beneficial effects:

[0176] 1. Explicitly overcomes the multi-hop communication bottleneck, laying the foundation for topology-aware scheduling.

[0177] By dynamically tracing the physical location of the predecessor through underlying scheduling and accumulating the shortest path transmission cost, the "far-end computing power trap" can be effectively avoided.

[0178] like Figure 3 As shown, with the number of edge servers increasing from 4 to 8, the complexity of multi-hop links in the physical network increases significantly. The complete algorithm of this invention (TA-BiRLH), leveraging a dual-graph joint perception mechanism, steadily increases its hit rate to 0.99 with increasing computing power, reaching a target value as low as -931. In contrast, the ablation benchmark (BiRLH) without the topology perception module, lacking a physical routing perspective, frequently misschedules data-intensive tasks to remote nodes, resulting in a hit rate that fluctuates wildly between 0.38 and 0.82. This experiment directly demonstrates the decisive role of "explicitly considering multi-hop forwarding costs" and "dual-graph joint state representation" in avoiding communication bottlenecks, ensuring the global rationality of the scheduling scheme in sparse topologies. Dual-graph perception improves decision-making quality: the enhanced graph encoder simultaneously captures task throughput requirements and edge link bandwidth heterogeneity, preventing agents from allocating high-data-volume tasks to congested nodes at the state space level, thus improving the global rationality of the scheduling scheme.

[0179] Figure 3 The impact of the number of edge servers on algorithm performance. Figure 3 Display a comparison of system target values ​​under different server sizes; Figure 4 The application hit rate comparison is shown. Data indicates that TA-BiRLH maintains an absolute lead across all scales, while the ablation benchmark suffers a performance precipitate due to topological blindness.

[0180] 2. Dual-map joint perception improves the quality of high-concurrency decision-making.

[0181] The enhanced graph encoder simultaneously captures task throughput requirements and edge link bandwidth heterogeneity, eliminating blind decision-making by the agent at the state space level.

[0182] like Figure 4 As shown, when the number of concurrent users increases to 20 and 25, the bottleneck of physical network communication begins to emerge. TA-BiRLH, by fusing dual-flow graph coding and cross-domain attention, simultaneously identifies task load distribution and link congestion nodes, maintaining the target value steadily at -686 and -440, with a hit rate of 0.55. In contrast, the baseline algorithm relying solely on task DAGs, unable to perceive physical routing costs, sees its target value deteriorate to a positive value (approximately 74), and its hit rate plummet to 0.22. This demonstrates that our method can accurately balance "remote computing power acceleration" and "local data transfer exemption benefits" in the fog of resource contention, maintaining scheduling robustness.

[0183] Figure 4The impact of the number of users on algorithm performance. Figure 5 This shows the trend of the target value as user load increases. Figure 6 For application hit rate comparison. Under high concurrency, the TA-BiRLH curve is flat, and the traditional algorithm suffers performance collapse due to the lack of modeling of communication penalty.

[0184] 3. Strong topology adaptability and real-world application capability.

[0185] The model learns the intrinsic mapping rule of "topological features - scheduling cost" and does not overfit to synthetic data structures.

[0186] like Figure 5 As shown, load testing was performed using a real industrial DAG from Alibaba Cluster Trace. The task topology exhibited highly irregular and long-tailed distribution characteristics. Traditional heuristic algorithms quickly fell behind due to rigid rules, with a hit rate only maintaining between 0.38 and 0.58. In contrast, TA-BiRLH achieved a perfect hit rate of 1.0 under light load with 10 / 15 users, and maintained a high hit rate of 0.77 even under severe congestion with 25 users. This demonstrates that the "logical DAG-physical topology" dual-graph mapping mechanism of this invention can effectively adapt to the multi-hop backhaul paths of real long-tailed distributed tasks, possessing extremely strong engineering deployment adaptability.

[0187] Figure 5 A bar chart comparing the performance of each algorithm in a real dataset scenario. Figure 7 A bar chart showing the target values ​​under actual load; Figure 8 The application hit rate is shown in the bar chart. TA-BiRLH achieved an overwhelming advantage across all load tiers, validating its potential for real-world deployment.

[0188] 4. Zero-shot generalization ability and learning of intrinsic optimization rules.

[0189] The lower-layer multi-hop routing simulation provides high-fidelity physical feedback to the upper layer. Combined with the PPO pruning mechanism and GAE advantage estimation, the agent can grasp the underlying scheduling rules rather than memorizing a fixed scale.

[0190] like Figure 6 As shown, the reinforcement learning agent was trained only in a fixed-scale environment with 20 users, and then directly tested across scales of 5 to 30 users. In the unseen high-pressure scenario of 30 users, the TA-BiRLH target value remained stable at -270, and the hit rate remained at 0.38; while the target value of the traditional CP algorithm deteriorated to 155, and the hit rate plummeted to a freezing point of 0.03. This demonstrates that the present invention successfully captures the inherent mapping relationship between latency, multi-hop communication, and computing resources in complex scheduling, and has the potential for zero-sample transfer to downstream ultra-large-scale environments without retraining.

[0191] Figure 6The zero-shot generalization performance of TA-BiRLH is shown for different user scales. Figure 9 Displays a comparison of target values ​​across different scales; Figure 10 The example demonstrates a cross-scale hit rate comparison. Even with mismatched training and test set sizes, TA-BiRLH maintains robust performance, validating its generalization ability.

[0192] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0193] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0194] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A multi-hop network topology-aware edge computing-dependent task scheduling method, characterized in that, include: S1, Obtain the directed acyclic graph of the task logic; Obtain the undirected graph of the edge physical network; initialize the parameters of the policy network, value network, and augmented graph encoder for the upper-layer near-end policy optimization; S2 inputs the task logic directed acyclic graph and the edge physical network undirected graph in parallel into the augmented graph encoder; bidirectional graph convolution is used to aggregate messages along the forward and reverse edges of the task logic directed acyclic graph, and the concatenation is used to obtain the task node-level feature vector; a graph attention network is used to extract link congestion and node connectivity features of the edge physical network undirected graph; and a cross-domain attention pooling mechanism is used to nonlinearly fuse the computational load features of the task logic directed acyclic graph and the topological bottleneck features of the edge physical network undirected graph to output the global environment state feature vector. S3, the actor network of the policy network receives the global environment state feature vector and outputs an action; the output action is to add a virtual scheduling constraint edge to the current task logic directed acyclic graph, which is used to force two subtasks that could originally be parallel to be transformed into a serial relationship with a sequential execution order. A dynamic action masking mechanism is used to ensure that adding virtual scheduling constraint edges does not disrupt the strict causal execution order of the original directed acyclic graph, thereby generating an enhanced logic graph. S4 receives the enhanced logic graph and the undirected graph of the edge physical network, performs multi-hop perceptual heuristic scheduling, and outputs a physical offloading scheme. S5, calculate the system multi-objective joint performance value including multi-hop forwarding cost based on the physical offloading scheme; define the immediate reward as the relative change between the system multi-objective joint performance value corresponding to the enhanced logic graph and the system multi-objective joint performance value before adding virtual scheduling constraint edges; Calculate the temporal difference residuals and estimate the cumulative advantage function using generalized advantage estimation; sample batch data from the empirical replay pool and update the actor network using the probability proportion pruning objective function optimized by the near-end policy, while simultaneously minimizing the mean square error to update the value network; repeat steps S2 to S5 until the policy network converges.

2. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 1, characterized in that, In step S1, the enhanced graph encoder adopts a dual-flow graph coding structure: The task graph encoding stream uses a bidirectional graph convolutional network, which includes a forward message passing layer and a backward message passing layer. The forward layer aggregates predecessor features along the direction from the start node to the end node, and the backward layer aggregates successor features along the direction from the end node to the start node. The bidirectional aggregation results are concatenated and then nonlinearly transformed to obtain task node-level features. The physical topology graph encoding stream uses a multi-head graph attention network to calculate the attention coefficients of each edge node with all its neighboring nodes, weighted aggregate the link bandwidth and computing resource characteristics of the neighboring nodes, and obtain the physical node-level features after multi-head splicing and averaging. In step S2, the cross-domain attention pooling uses task node-level features as query vectors and physical node-level features as key-value vectors to calculate cross-domain attention weights and perform weighted pooling on physical features. Then, it is concatenated with the global pooling features of the task graph and mapped through a fully connected layer to output a global environment state feature vector of fixed dimensions.

3. The multi-hop network topology-aware edge computing-dependent task scheduling method according to claim 1, characterized in that, In step S3, the dynamic action masking mechanism imposes three rigid constraints on the candidate virtual edges: the non-repetition constraint requires that the candidate edge has a value of zero in the current graph adjacency matrix; the non-redundancy constraint requires that the reachability of the candidate edge from the start node to the end node in the current graph transitive closure matrix is ​​zero; and the strict topological order constraint requires that the index of the start node in the current graph topological sort is less than the index of the end node. The output action space of the actor network is the probability distribution on the set of legal candidate edges that satisfy all the above constraints. The actual virtual edges added are determined by sampling.

4. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 1, characterized in that, In step S4, the enhanced logic graph and the undirected graph of the edge physical network are received, multi-hop awareness heuristic scheduling is performed, and the output physical offloading scheme includes: Maintain a global ready task queue and push all task nodes with an in-degree of zero into the queue; Calculate the underlying priority value of each task in the ready task queue and rearrange them in descending order; Pop up the first task in the queue and filter the set of candidate physical nodes that meet the constraints of the operating environment; For each candidate physical node, trace the set of physical execution locations of all direct predecessor tasks of this task, call the shortest path algorithm to query the shortest route path from the physical execution location of the predecessor task to the current candidate physical node in the undirected graph of the edge physical network, accumulate the transmission delay of each link on the shortest route path based on the link communication rate to calculate the multi-hop forwarding cost, and combine the local computing delay, queuing delay and energy consumption cost of the current candidate physical node to predict the incremental cost of the scheduling decision on the system's comprehensive objective function; Select the physical node with the lowest incremental cost to bind the current task, update the idle timestamp of the physical node, push the subsequent tasks that meet the dependency conditions into the ready task queue, and execute in a loop until all tasks are scheduled and output the physical unloading solution.

5. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 4, characterized in that, In step S4, the multi-hop forwarding cost is calculated as follows: when the bound physical node of the direct predecessor task of the task to be scheduled is the same as the current candidate node, the cost is zero; when they are different nodes, the shortest path algorithm is called to query the shortest route from the predecessor node to the candidate node in the physical graph. The multi-hop forwarding cost is equal to the sum of the transmission delays obtained by dividing the output data of all links on the path by the corresponding communication rate.

6. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 4, characterized in that, In step S4, the incremental cost is the weighted sum of multi-hop forwarding cost, local computation latency, queuing latency, and energy consumption cost; local computation latency is equal to the task computation amount divided by the candidate node's computation capacity. Queuing delay is equal to the positive part of the difference between the current idle timestamp of the candidate node and the latest completion time of all direct predecessor tasks of the task; Energy consumption cost equals the unit energy consumption coefficient of the candidate node multiplied by the task computation amount; each item is multiplied by its corresponding weight coefficient and then summed to obtain the incremental cost predicted for the candidate node.

7. The multi-hop network topology-aware edge computing-dependent task scheduling method according to claim 1, characterized in that, In step S5, the system multi-objective joint performance value is a weighted combination of total processing latency, total system energy consumption, total execution cost and application hit rate. The total processing latency is the sum of the computation latency of all tasks and the multi-hop forwarding latency of all cross-node dependency edges. The application hit rate is the proportion of the number of applications that successfully complete all sub-task scheduling within the preset deadline to the total number of applications. The instant reward is the difference between the efficiency value of the enhanced graph and the efficiency value before adding the edge. The reward is positive when the efficiency value is better after enhancement, and negative otherwise.

8. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 1, characterized in that, In step S5, the objective function is: calculate the probability ratio of the new strategy and the old strategy in the sampling action, trim the probability ratio to a preset interval, take the product of the trimmed ratio and the dominance function, and then take the minimum value of the product of the untrimmed ratio and the dominance function as the optimization objective; the value network is updated by minimizing the mean square error of the current value network output by minimizing the temporal difference objective value; the dominance estimation adopts generalized dominance estimation, which is a weighted sum of multiple temporal difference errors and the cumulative dominance function.

9. The multi-hop network topology-aware edge computing dependent task scheduling method according to claim 1, characterized in that, The nodes of the directed acyclic graph of the task logic include computational and data volume attributes, task type labels, and deadline constraints. The edge attributes of the undirected graph of the edge physical network include link communication rate, link latency, and reliability metrics. The enhanced graph encoder parameters are initialized uniformly using Xavier; The parameters of the actor network and value network are orthogonally initialized; The experience replay pool adopts priority experience replay, and the sampling priority is allocated according to the size of the time difference error. The larger the error, the higher the probability of the sample being sampled.

10. A multi-hop network topology-aware edge computing-dependent task scheduling system, characterized in that, include: The dual-graph construction module is used to obtain the task logic directed acyclic graph and the edge physical network undirected graph, and to perform structured storage of the node attributes of the task logic directed acyclic graph and the node and edge attributes of the edge physical network undirected graph. The state encoding module includes an enhanced graph encoder, which encodes the task logic directed acyclic graph and the edge physical network undirected graph in parallel. It extracts task graph features and physical topology graph features through bidirectional graph convolution and graph attention network, respectively, and fuses them through cross-domain attention pooling mechanism to output a global environment state feature vector. The policy sampling module includes an actor network and a value network for the near-end policy optimization policy network. The actor network is used to receive the global environment state feature vector and output the actions with added virtual scheduling constraint edges. The value network is used to evaluate the value of the current state. The policy sampling module also includes a dynamic action mask unit, which is used to verify the legality of the actions output by the actor network. The multi-hop scheduling module is used to receive the augmented logic graph and the undirected graph of the edge physical network, and perform multi-hop awareness heuristic scheduling. The multi-hop scheduling module includes a ready queue management unit, a low-level priority calculation unit, a predecessor position tracing unit, a shortest path query unit, a multi-hop forwarding cost calculation unit, and an incremental cost evaluation unit, which are used to bind each task in the augmented logic graph to the physical node that minimizes the incremental cost and output a physical offloading scheme. The reward update module is used to calculate the joint performance value of the system with multi-hop forwarding costs based on the physical offloading scheme, calculate the immediate reward, estimate the cumulative advantage function through generalized advantage, and update the actor network and value network using the probabilistic pruning objective function optimized by the near-end policy.

Images