Edge computing node task unloading method based on fruit fly optimization and ant colony collaboration

Abstract

The invention discloses an edge computing node task unloading method based on fruit fly optimization and ant colony collaboration. The method comprises the following steps: S1, obtaining a real-time state data set of all edge computing nodes in an edge computing environment through a preset monitoring module; s2, performing feature data extraction on the to-be-unloaded task based on the state data set; s3, performing global search on the state data set and the task feature data by using a fruit fly optimization algorithm to generate a preliminary task unloading candidate node set; s4, for the initial task unloading candidate node set, constructing a connected graph between edge computing nodes by adopting an ant colony collaborative algorithm, and locally optimizing an unloading path between candidate nodes; and S5, distributing and migrating the task to be unloaded according to the determined unloading path and the target node. According to the method, the communication delay and bandwidth between the nodes are considered, and dynamic weighting is carried out on the resource remaining and load state of the target node, so that the accuracy and flexibility of unloading path construction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and particularly to a task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation. Background Art

[0002] With the rapid development of Internet of Things and 5G communication technologies, edge computing, as a distributed computing architecture close to end-users, shows broad prospects in reducing data transmission latency, reducing the load of the central cloud, and enhancing real-time response capabilities. Especially in application scenarios such as intelligent manufacturing, autonomous driving, and remote healthcare that have high requirements for timeliness and computing performance, the edge offloading of tasks has become a key technology. However, how to efficiently perform task offloading scheduling in an edge computing environment with multi-node and multi-task dynamic changes remains a difficult point in current research and practice.

[0003] Existing edge computing task offloading methods mainly include heuristic methods based on static rules and scheduling strategies based on single optimization algorithms. Heuristic methods such as "shortest latency first" and "maximum resource node first" are simple to implement, but often ignore the dynamic coupling relationship between task characteristics and node states, which is likely to cause partial node overload or task latency timeout while the resource utilization rate is limited. The strategies based on single optimization algorithms, although having certain global search or local convergence properties, are prone to problems such as slow convergence speed, unstable scheduling paths, and lack of robustness of scheduling results when facing strong node heterogeneity, high task diversity, and unstable communication environments in edge computing networks.

[0004] Especially in the edge computing scenario with multi-task concurrency, existing technologies often have difficulty in simultaneously considering the computational matching, communication cost, and network connectivity of task offloading. The scheduling results rely too much on static resource information, ignoring the real-time nature, computational complexity, and dependencies between tasks, resulting in unreasonable offloading path selection and unbalanced task allocation. In addition, many existing offloading methods lack a systematic path optimization mechanism during task scheduling and cannot effectively evaluate and dynamically adjust the migration paths of tasks between multiple nodes, thus affecting the overall scheduling performance and service quality.

[0005] In summary, existing edge computing task offloading technologies have obvious deficiencies in task feature recognition, node performance evaluation, offloading path selection, and scheduling strategy stability, and are difficult to meet the intelligent offloading requirements in a multi-node, multi-task, and highly dynamic environment. Therefore, there is an urgent need for a task offloading method that can integrate multi-source state information and task characteristics, and at the same time has global search capabilities and a path optimization mechanism to improve offloading efficiency and the overall performance of the system. Summary of the Invention

[0006] An object of the present invention is to propose a task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation. The present invention not only considers the communication delay and bandwidth between nodes, but also dynamically weights the remaining resources and load status of the target node, thereby significantly improving the accuracy and flexibility of the offloading path construction.

[0007] A task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation according to an embodiment of the present invention includes the following steps:

[0008] S1. Obtain the real-time status data set of all edge computing nodes in the edge computing environment through a preset monitoring module;

[0009] S2. Extract feature data of the task to be offloaded based on the status data set;

[0010] S3. Use the fruit fly optimization algorithm to globally search the status data set and the task feature data to generate a preliminary task offloading candidate node set, and the preliminary task offloading candidate node set reflects the comprehensive performance of each edge computing node under the condition of meeting the task offloading requirements;

[0011] S4. For the preliminary task offloading candidate node set, use the ant colony cooperation algorithm to construct a connected graph between edge computing nodes, and locally optimize the offloading path between candidate nodes. The optimization result is used as the basis for determining the task offloading path and the target node;

[0012] S5. Execute the task offloading scheduling operation according to the task offloading path and the target node, and allocate and migrate the task to be offloaded according to the determined offloading path and the target node.

[0013] Optionally, S1 includes the following steps:

[0014] S11. Set the edge computing node set as N;

[0015] S12. Collect the remaining computing resources, current task load, available uplink network bandwidth, and average communication delay of each edge computing node n i to construct a single-node status vector:

[0016] S i =(R i , L i , B i , D i );

[0017] wherein, R i represents the remaining computing resources of the i-th edge computing node, L i represents the current task load of the i-th edge computing node, which is defined as the CPU occupancy rate of the currently assigned task, and the value range is [0, 1], Bi represents the available uplink network bandwidth of the i-th edge computing node, D i represents the average communication delay between the i-th edge computing node and the offloading initiating node;

[0018] S13. Integrate the state vectors of all edge computing nodes to construct a system-level real-time state dataset of edge computing nodes:

[0019] S = {S i | i = 1, 2,..., M};

[0020] S14. Perform standardized preprocessing on the real-time state dataset S, normalize each dimensional attribute therein to a unified dimension, and generate a standardized real-time state dataset S ′ .

[0021] Optionally, the S2 includes the following steps:

[0022] S21. Set the set of tasks to be offloaded as T = {t1, t2,..., t K}, where t j represents the j-th task to be offloaded, and K is the total number of tasks to be offloaded;

[0023] S22. For each task to be offloaded t j , collect the task data volume Q j , computational complexity C j , real-time requirement τ j and task dependency δ j , and construct a single-task feature vector T j ;

[0024] S23. Integrate the single-task feature vectors of all tasks to be offloaded to construct a task feature dataset T feat .

[0025] Optionally, the S3 includes the following steps:

[0026] S31. Take the standardized real-time state dataset and the task feature dataset as inputs, and initialize the fruit fly population set P = {p1, p2,..., p N}, where p k represents the k-th fruit fly individual, and each individual p k corresponds to an offloading matching scheme of tasks to edge computing nodes;

[0027] S32. Introduce a task-aware dynamic search radius control mechanism. The dynamic search radius control mechanism reflects that the more urgent the task and the more congested the node, the smaller the search step size, so that the search focuses on the vicinity of low-latency and high-resource nodes. Define the fruit fly individual p kThe search step size in each round of search iteration is as follows:

[0028]

[0029] where is the fruit fly individual p k when processing task t j The search step size, indicating the local search range of the fruit fly individual p k in the offloading node space. λ0 is the initially set search reference step size, defining the average jump amplitude of the fruit fly at the initial stage of task offloading optimization. τ j is the real-time requirement of task t j indicating the maximum allowable response delay of the task. τ max is the maximum allowable delay threshold set by the system. θ is the node load perception adjustment coefficient, controlling the coupling strength between the task urgency and the node congestion state. L′ i(k,j) is the normalized task load of the edge computing node i(k,j), representing the relative saturation of the node in processing the current task. R′ i(k,j) is the normalized remaining computing resource of the edge computing node i(k,j), representing the relative computing ability of the node available for new task processing;

[0030] S33. Construct a task-aware multi-objective comprehensive performance function. The task-aware multi-objective comprehensive performance function strongly couples task characteristics with node states, reflecting the combined effects of transmission bottlenecks, computing power matching, and load pressure during the task offloading process, and evaluates the offloading scheme of the fruit fly individual p k for task t j :

[0031]

[0032] where F k,j represents the performance cost value of the selected node of the fruit fly individual p k for task t j . The smaller the value, the better the scheme. R′ i(k,j) , L′ i(k,j) , B′ i(k,j) , D′ i(k,j) are respectively the normalized remaining computing resource, the current task load, the network bandwidth, and the communication delay of the selected node of the fruit fly individual p k for task t j . w1, w2, w3, w4 are the multi-objective weight factors set by the system;

[0033] S34. Sort the performance cost values F k,j of all individuals, and select the edge computing node number with the minimum cost for each task t j among all individuals

[0034]

[0035] S35. All tasks t j The corresponding optimal edge computing node number Perform integration and build a preliminary set C of candidate nodes for task offloading.

[0036] Optionally, S4 includes the following steps:

[0037] S41. The nodes in the preliminary task offloading candidate node set C constitute a node set V in the task offloading connectivity graph structure, and construct a task offloading connectivity graph G = (V, E), where E is a set of communication paths between candidate nodes with the possibility of direct offloading, reflecting the network connectivity between the nodes;

[0038] S42. For any two candidate nodes n in the task offloading connected graph G i ,n j , a dynamic offloading cost function is introduced to quantify the comprehensive performance of the task offloading path between nodes:

[0039]

[0040] Among them, D i,j Represents node n i To node n j The average communication delay of B i,j represents the available transmission bandwidth between two nodes, L′ j For node n j The normalized current task load, R′ j For node n j The normalized remaining computational resources, Δ i,j It represents the real-time cost fluctuation factor caused by network fluctuation and dynamic change of node status. μ1, μ2, μ3 and μ4 are the weight coefficients preset by the system.

[0041] S43. Initialize the pheromone matrix τ = {τ i,j ∣n i ,n j ∈V}, and define the heuristic factor matrix η according to the dynamic offloading cost function. The η in the heuristic factor matrix i,j Directly reflects from node n i to n j The attractiveness of the unloading path relative to the cost function;

[0042] S44. Define the dynamic adaptation factor ζ i,jTo enhance the selection probability of resource - dominant nodes and adapt to the heterogeneity of each node in the edge - computing environment, the dynamic adaptation factor enables that when node n j has abundant resources and low load, the probability of being selected as an offloading target is significantly increased, thus alleviating local resource tension:

[0043]

[0044] where ξ is the resource - and - load sensitivity adjustment coefficient;

[0045] S45. Introduce each artificial ant into the task - offloading connectivity graph G to construct a task - offloading path. The k - th ant starts from the offloading - initiating node and, in the set of candidate nodes that have not been visited calculate the dynamic transfer probability of transferring from node n i to node n j . The dynamic transfer probability comprehensively reflects the historical goodness of the offloading path, the current expected cost, and the dynamic advantages of node resources, thus promoting the global and local collaborative optimization of the ant - colony algorithm in edge - computing task offloading:

[0046]

[0047] where α controls the influence degree of pheromone, β controls the role of the heuristic factor, and γ controls the importance of the dynamic adaptation factor ζ i,j ;

[0048] S46. After each ant completes the construction of an offloading path, calculate the total offloading cost:

[0049]

[0050] where Ψ k is the cumulative cost of the path constructed by the k - th ant;

[0051] S47. Update the pheromone matrix according to the total offloading costs of the paths constructed by each ant:

[0052]

[0053] where ρ is the pheromone evaporation coefficient, N a is the total number of ants participating in path construction, and Q is the pheromone - release intensity constant;

[0054] S48. Repeat steps S44 to S47 within the preset number of iterations until the convergence condition is reached;

[0055] Finally, select the path with the lowest score from all the constructed offloading paths, and the terminal node of the path with the lowest score is determined as the target node for the final task offloading.

[0056] Optionally, S5 includes the following steps:

[0057] S61. Output the offloading scheduling result according to the finally determined task offloading path and the target node. The task offloading path is the path Path with the minimum total cost in the constructed path set, * and the target node is the end node of this path.

[0058] S62. For each task t in the task set T to be offloaded, j allocate and migrate it according to the corresponding offloading path and the target node; perform allocation and migration.

[0059] S63. During the task migration process, use the multi-hop forwarding mechanism to orderly transmit the task data along the path to the target node, and perform task data integrity verification and reception confirmation at each hop;

[0060] S64. When the target node successfully receives the task t, j initialize the task execution environment according to the task feature data T j and the state vector of this node to complete the offloading task deployment;

[0061] S65. Record the migration completion time, startup delay and execution duration of the task t j in real time, and compare them with the task real-time constraint τ j to generate task offloading evaluation metrics;

[0062] S66. If the offloading fails due to path interruption, node abnormality or insufficient resources, trigger the backoff scheduling mechanism, reselect the sub-optimal offloading path and the target node from the candidate node set, and perform task rescheduling and migration operations.

[0063] The beneficial effects of the present invention are as follows:

[0064] (1) By constructing a task-aware dynamic search radius control mechanism, the present invention realizes dynamically adjusting the search step size during the global search of fruit flies. Specifically, the system dynamically calculates the search step size according to the real-time requirements of the task and the resource load status of the candidate nodes, so that the more urgent the task and the more congested the node, the more the search approaches the nodes with high resources and low latency, thereby improving the effectiveness of offloading matching and effectively avoiding the problems of blind expansion of the search space or falling into local optimality.

[0065] (2) The present invention constructs a task-aware multi-objective comprehensive performance function, comprehensively considering multiple factors such as the data volume, computational complexity, communication delay, resource availability, and load condition of the nodes, and introducing a weight adjustment mechanism to achieve strong coupling between task characteristics and node states, effectively breaking the limitation of the traditional offloading model with a single optimization target of delay or bandwidth, and being able to more comprehensively evaluate the comprehensive carrying capacity of candidate nodes in the offloading decision.

[0066] (3) Based on the initially selected set of offloading nodes, the present invention uses the ant colony algorithm to locally optimize the offloading path, and introduces a dynamic offloading cost function, a heuristic factor matrix, and a resource-aware dynamic adaptation factor, not only considering the communication delay and bandwidth between nodes, but also dynamically weighting the remaining resources and load status of the target nodes, thereby significantly improving the accuracy and flexibility of the offloading path construction. Description of the Drawings

[0067] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0068] Figure 1 is a flowchart of a method for offloading tasks of edge computing nodes based on fruit fly optimization and ant colony cooperation proposed by the present invention. Detailed Embodiments

[0069] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0070] Refer to Figure 1 , a method for offloading tasks of edge computing nodes based on fruit fly optimization and ant colony cooperation, includes the following steps:

[0071] S1. Obtain the real-time status data set of all edge computing nodes in the edge computing environment through a preset monitoring module;

[0072] S2. Extract feature data of the task to be offloaded based on the status data set;

[0073] S3. Use the fruit fly optimization algorithm to globally search the status data set and the task feature data to generate a preliminary set of candidate nodes for task offloading, and the preliminary set of candidate nodes for task offloading reflects the comprehensive performance of each edge computing node under the condition of meeting the task offloading requirements;

[0074] S4. For the initially obtained set of candidate nodes for task offloading, use the ant colony cooperation algorithm to construct a connected graph among edge computing nodes, and locally optimize the offloading paths between candidate nodes. The optimization results are used as the basis for determining the task offloading path and the target node;

[0075] S5. Execute the task offloading scheduling operation according to the task offloading path and the target node, and allocate and migrate the tasks to be offloaded according to the determined offloading path and target node.

[0076] In this embodiment, S1 includes the following steps:

[0077] S11. Set the set of edge computing nodes as N;

[0078] S12. Collect the remaining computing resources, current task load, available uplink network bandwidth, and average communication delay of each edge computing node n i to construct a single-node state vector:

[0079] S i =(R i ,L i ,B i ,D i );

[0080] where R i represents the remaining computing resources of the i-th edge computing node, L i represents the current task load of the i-th edge computing node, defined as the CPU occupancy rate of the currently assigned tasks, with a value range of [0,1], B i represents the available uplink network bandwidth of the i-th edge computing node, and D i represents the average communication delay between the i-th edge computing node and the offloading initiating node;

[0081] S13. Integrate the state vectors of all edge computing nodes to construct a system-level real-time state dataset of edge computing nodes:

[0082] S={S i |i = 1,2,...,M};

[0083] S14. Perform standardized preprocessing on the real-time state dataset S, normalize each dimensional attribute in it to a unified dimension, and generate a standardized real-time state dataset S ′ .

[0084] In this embodiment, S2 includes the following steps:

[0085] S21. Set the set of tasks to be offloaded as T={t1,t2,…,t K}, where t jRepresents the jth task to be offloaded, and K is the total number of tasks to be offloaded;

[0086] S22. For each task t to be offloaded j , collect the task data volume Q j , computing complexity C j , real-time requirement τ j and task dependency δ j , and construct a single-task feature vector T j ;

[0087] S23. Integrate the single-task feature vectors of all tasks to be offloaded to construct a task feature dataset T feat .

[0088] In this embodiment, S3 includes the following steps:

[0089] S31. Use the standardized real-time status dataset and the task feature dataset as inputs to initialize the fruit fly population set P = {p1, p2,..., p N}), where p k represents the kth fruit fly individual, and each individual p k corresponds to an offloading matching scheme from tasks to edge computing nodes;

[0090] S32. Introduce a task-aware dynamic search radius control mechanism. The dynamic search radius control mechanism reflects that the more urgent the task and the more congested the node, the smaller the search step size, so that the search focuses on the vicinity of low-latency and high-resource nodes. Define the search step size of the fruit fly individual p k in each round of search iteration as:

[0091]

[0092] where is the search step size of the fruit fly individual p k when processing task t j , representing the local search range of the fruit fly individual p k in the offloading node space. λ0 is the initially set search reference step size, defining the average jump amplitude of the fruit fly at the initial stage of task offloading optimization. τ j is the real-time requirement of task t j , representing the maximum allowable response delay of the task. τ max is the maximum allowable delay threshold set by the system. θ is the node load perception adjustment coefficient, controlling the coupling strength between the task urgency and the node congestion state. L′ i(k,j) is the standardized task load of edge computing node i(k, j), representing the relative saturation of the node in processing tasks currently. R′ i(k,j)is the normalized remaining computing resource of edge computing node i(k,j), indicating the relative computing power of the node that can be used to process new tasks;

[0093] S33. Construct a task-aware multi-objective comprehensive performance function. The task-aware multi-objective comprehensive performance function strongly couples the task characteristics with the node status, reflects the comprehensive effects of transmission bottlenecks, computing power matching, and load pressure during task offloading, and performs a multi-objective comprehensive performance function on the individual p of Drosophila. k For task t j The uninstallation scenarios are evaluated:

[0094]

[0095] Among them, F k,j represents the fruit fly individual p k For task t j The performance cost of the selected node. The smaller the value, the better the solution. R′ i(k,j) , L′ i(k,j) , B′ i(k,j) , D′ i(k,j) The fruit fly individual p k For task t j The standardized remaining computing resources, current task load, network bandwidth and communication delay of the selected node. w1, w2, w3, w4 are the multi-objective weight factors set by the system.

[0096] S34. Performance cost F for all individuals k,j Sort and select each task t j The edge computing node number with the minimum cost among all individuals

[0097]

[0098] S35. All tasks t j The corresponding optimal edge computing node number Perform integration and build a preliminary set C of candidate nodes for task offloading.

[0099] In this implementation, S4 includes the following steps:

[0100] S41. The nodes in the preliminary task offloading candidate node set C constitute a node set V in the task offloading connectivity graph structure, and construct a task offloading connectivity graph G = (V, E), where E is a set of communication paths between candidate nodes with the possibility of direct offloading, reflecting the network connectivity between the nodes;

[0101] S42. For any two candidate nodes n in the task offloading connected graph G i ,n j, a dynamic offloading cost function is introduced to quantify the comprehensive performance of the task offloading paths between nodes:

[0102]

[0103] Among them, D i,j represents the average communication delay from node n i to node n j , B i,j represents the available transmission bandwidth between two nodes, L′ j is the normalized current task load of node n j , R′ j is the normalized remaining computing resource of node n j , Δ i,j represents the real-time cost fluctuation factor caused by network fluctuations and dynamic changes in node states, and μ1, μ2, μ3, μ4 are preset weight coefficients of the system;

[0104] S43. Initialize the pheromone matrix τ = {τ i,j ∣n i , n j ∈V}, and define the heuristic factor matrix η according to the dynamic offloading cost function. The η i,j in the heuristic factor matrix directly reflects the attractiveness of the offloading path from node n i to n j relative to the cost function;

[0105] S44. Define the dynamic adaptation factor ζ i,j to strengthen the selection probability of resource-advantage nodes and adapt to the heterogeneity of each node in the edge computing environment. The dynamic adaptation factor makes the probability of being selected as an offloading target significantly increase when the resources of node n j are abundant and the load is low, thereby alleviating local resource tension:

[0106]

[0107] Among them, ξ is the resource and load sensitive adjustment coefficient;

[0108] S45. Introduce each artificial ant into the task offloading connected graph G to construct the task offloading path. The kth ant starts from the offloading initiating node and, in the set of candidate nodes that have not been visited , calculates the dynamic transition probability of transferring from node n i to node n j . The dynamic transition probability comprehensively reflects the historical goodness, current expected cost, and dynamic resource advantages of the offloading path, thereby promoting the global and local collaborative optimization of the ant colony algorithm in edge computing task offloading:

[0109]

[0110] Among them, α controls the influence degree of pheromone, β controls the role of heuristic factor, and γ controls the importance of the dynamic adaptation factor ζ i,j ;

[0111] S46. When each ant completes the construction of an offloading path, calculate the total offloading cost:

[0112]

[0113] Among them, Ψ k is the cumulative cost of the path constructed by the k-th ant;

[0114] S47. According to the total offloading costs of the paths constructed by each ant, update the pheromone matrix:

[0115]

[0116] Among them, ρ is the pheromone evaporation coefficient, N a is the total number of ants participating in path construction, and Q is the pheromone release intensity constant;

[0117] S48. Repeat steps S44 to S47 within the preset number of iterations until the convergence condition is reached;

[0118] Finally, select the path with the lowest score from all the constructed offloading paths, and the terminal node of the path with the lowest score is determined as the target node for the final task offloading.

[0119] In this embodiment, S5 includes the following steps:

[0120] S61. Output the offloading scheduling result according to the finally determined task offloading path and the target node. The task offloading path is the path Path with the minimum total cost in the constructed path set * , and the target node is the end node of this path

[0121] S62. For each task t in the task set T to be offloaded j allocate and migrate it according to the corresponding offloading path and the target node ;

[0122] S63. During the task migration process, use the multi-hop forwarding mechanism to orderly transmit the task data along the path to the target node, and perform task data integrity verification and reception confirmation at each hop;

[0123] S64. When the target node successfully receives the task t j after that, according to the task characteristic data T jWith the state vector of this node Initialize the task execution environment and complete the deployment of the offloading task;

[0124] S65. For task t j Real-time record the migration completion time, start-up delay, and execution duration, and compare them with the task real-time constraint τ j to generate task offloading evaluation metrics;

[0125] S66. If the offloading fails due to path interruption, node abnormality, or resource shortage, trigger the backoff scheduling mechanism, reselect the sub-optimal offloading path and target node from the candidate node set, and perform task rescheduling and migration operations.

[0126] Example 1:

[0127] At 9:10 am on November 16, 2024, in a large intelligent logistics and warehousing center located in City A, during the operation of daily high-load tasks such as logistics identification, path planning, and video analysis, the edge computing system showed abnormal phenomena such as task backlog and increased response delay. The technical team found that due to the sharp increase in the number of logistics vehicles entering the warehouse that morning, the image recognition and dynamic path calculation tasks were densely triggered within a certain period of time, and the system scheduling module could not evenly distribute the computing pressure in a timely manner, resulting in processing bottlenecks in individual edge nodes such as Node_07 and Node_09.

[0128] At 9:12 am on the same day, the edge control master node "192.168.0.1" detected the latest status data of each edge computing node through the deployed real-time monitoring module, showing that the CPU utilization rate of node Node_07 (IP is 192.168.0.107) had continued to be higher than 92%, the uplink bandwidth had dropped to 18.3 Mbps, and the average communication delay had reached 14.7 ms. At the same time, node Node_03 (IP is 192.168.0.103) was in good condition, with a CPU idle rate of 63%, a delay of only 4.3 ms, and a stable network bandwidth of 38 Mbps.

[0129] At this time, an image segmentation task numbered T_202411160934 was triggered from the high-definition camera on the east side of the warehouse. The task data volume was 44.7 MB, the estimated computational complexity was 7.6 GFLOPs, and the set real-time threshold for the task was 65 ms. This task was originally scheduled to be processed by the physically adjacent node Node_07, but due to its resource overload, the offloading scheduling mechanism needed to be triggered immediately.

[0130] The system calls the task offloading method proposed by the present invention. First, it inputs the task feature vector of T_202411160934 {Q = 44.7, C = 7.6, τ = 65, δ = no dependency} into the scheduling module, and at the same time retrieves the standardized status data of each node. The fruit fly optimization module initializes the population size to 80 individuals, and each individual represents a node task allocation scheme. The task real-time performance and the target node resource status are dynamically coupled to form the search radius λ. The search step size λ of the current task is calculated as 0.37, and the fruit fly search focuses near the nodes with healthy status of Node_02, Node_03, and Node_05.

[0131] After the 12th generation of search iteration, the system outputs the offloading candidate node set as:

[0132] Node_02(192.168.0.102); Node_03(192.168.0.103); Node_05(192.168.0.105).

[0133] Among them, Node_03 has the lowest score (0.847) in the multi-objective performance function evaluation and enters the next path optimization.

[0134] The system then constructs an offloading connection graph based on this candidate node set and introduces a dynamic offloading cost function. In the path evaluation, the communication path delay from Node_07 to Node_03 is 4.1 ms, the link bandwidth is 35.2 Mbps, the current CPU idle of node Node_03 is 4.5 GHz, and the load is 38%. The system dynamic offloading cost ψ is evaluated as 1.93.

[0135] The ant colony optimization module initializes 15 artificial ants for path search. The heuristic factor η and the dynamic adaptation factor are calculated according to the real-time node resources. Finally, all paths are constructed in the 18th round, and the system selects the path Node_07→Node_03 with the minimum total cost as the final offloading path.

[0136] At 9:13:04, the system initiates an offloading operation. The task data is packed and sent to the target node Node_03 through the UDP protocol, which is divided into 37 sub-packets in total, and the size of each sub-packet is about 1.2 MB. The three-way handshake mechanism is used for data integrity verification.

[0137] At 9:13:06, Node_03 successfully receives all task data packets and sends an acknowledgment signal. At the same time, it pre-configures the task running environment according to its own resource status and starts the "image boundary segmentation" module in the container.

[0138] At 9:13:09, the task execution was completed. The system recorded the total task response time as 43 ms, which is far lower than the original set threshold of 65 ms, meeting the system's real-time requirements. The data output after task processing was automatically connected to the logistics scheduling module, completing the migration of this job node.

[0139] To verify the performance superiority of the present invention in a real scenario, system technicians compared and tested the performance of the traditional greedy offloading method and the method of the present invention within three days from November 14th to November 16th. Using real scheduling data (about 4,800 tasks and 12 nodes), the test environment parameters are as follows:

[0140]

[0141] During the peak offloading period (10:00 - 11:30 on November 16th), the system also extracted task log samples for item-by-item analysis:

[0142]

[0143] As can be seen above, by adopting the offloading scheme of the present invention, not only the risk of node load overload is successfully avoided, but also the offloading path is optimized, enabling the task response time to be stably maintained between 40 and 50 ms, ensuring the operation efficiency and security of the system under high concurrency. In contrast, the traditional strategy shows serious response lags when nodes are not unloaded in a timely manner, and even causes some tasks to fail to be processed.

[0144] In summary, the embodiment fully demonstrates the application process and technical advantages of the present invention in actual industrial scenarios, proving its offloading efficiency, robustness, and decision-making intelligence in dealing with high-dynamic edge computing environments, and having significant engineering practical value.

[0145] The present invention realizes the dynamic adjustment of the search step size during the global search of fruit flies by constructing a task-aware dynamic search radius control mechanism. Specifically, the system dynamically calculates the search step size according to the real-time requirements of tasks and the resource load status of candidate nodes, so that the more urgent the task and the more congested the node, the search approaches nodes with high resources and low latency, thereby improving the effectiveness of offloading matching and effectively avoiding the problems of blind expansion of the search space or falling into local optima.

[0146] The present invention constructs a task-aware multi-objective comprehensive performance function, comprehensively considering multiple factors such as the data volume, computational complexity, communication delay, resource availability, and load of nodes of tasks, and introducing a weight adjustment mechanism to achieve strong coupling between task characteristics and node states, effectively breaking the limitations of the traditional offloading model with a single optimization objective of delay or bandwidth, and being able to more comprehensively evaluate the comprehensive bearing capacity of candidate nodes in offloading decisions.

[0147] Based on the initially selected set of offloading nodes, the present invention uses the ant colony algorithm to locally optimize the offloading path, and introduces a dynamic offloading cost function, a heuristic factor matrix, and a resource-aware dynamic adaptation factor. It not only considers the communication delay and bandwidth between nodes, but also dynamically weights the remaining resources and load status of the target node, thereby significantly improving the accuracy and flexibility of offloading path construction.

[0148] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation, characterized in that, It includes the following steps: S1. Obtain the real-time status data set of all edge computing nodes in the edge computing environment through a preset monitoring module; S2. Extract feature data for the tasks to be offloaded based on the status data set; S3. Use the fruit fly optimization algorithm to perform a global search on the status data set and the task feature data, generate a preliminary task offloading candidate node set, and the preliminary task offloading candidate node set reflects the comprehensive performance of each edge computing node under the condition of meeting the task offloading requirements; S4. For the preliminary task offloading candidate node set, use the ant colony cooperation algorithm to construct a connectivity graph among edge computing nodes, and perform local optimization on the offloading paths between candidate nodes. The optimization result is used as the basis for determining the task offloading path and the target node; S5. Execute the task offloading scheduling operation according to the task offloading path and the target node, and allocate and migrate the tasks to be offloaded according to the determined offloading path and the target node.

2. The task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation according to claim 1, wherein The S1 includes the following steps: S11. Set the edge computing node set as N; S12. Collect each edge computing node n i remaining computing resources, current task load, available uplink network bandwidth, and average communication delay to construct a single-node state vector: S i = (R i , L i , B i , D i ); Among them, R i represents the remaining computing resources of the i-th edge computing node, L i represents the current task load of the i-th edge computing node, defined as the CPU occupancy rate of the currently assigned tasks, with a value range of [0, 1], B i represents the available uplink network bandwidth of the i-th edge computing node, D i represents the average communication delay between the i-th edge computing node and the offloading initiation node; S13. Integrate the status vectors of all edge computing nodes to construct a system-level edge computing node real-time status data set: S = {S i | i = 1, 2, ..., M}; S14. Perform standardized preprocessing on the real-time status data set S, normalize each dimensional attribute therein to a unified dimension, and generate a standardized real-time status data set S ′ .

3. A task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation according to claim 1, characterized in that The S2 includes the following steps: S21. Set the set of tasks to be offloaded as T = {t1, t2, …, t K}, where t j represents the j-th task to be offloaded, and K is the total number of tasks to be offloaded; S22. For each task t to be unloaded j , collect the task data volume Q j , the computational complexity C j , the real-time requirement τ j , and the task dependency δ j , and construct a single-task feature vector T j ; S23. Integrate the single-task feature vectors of all tasks to be unloaded to construct a task feature dataset T feat .

4. A task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation according to claim 1, characterized in that The S3 includes the following steps: S31. Take the standardized real-time status data set and the task feature data set as inputs, and initialize the fruit fly population set P = {p1, p2,..., p N}, where p k represents the k-th fruit fly individual, and each individual p k corresponds to an offloading matching scheme from a task to an edge computing node; S32. Introduce a task-aware dynamic search radius control mechanism. The dynamic search radius control mechanism reflects that the more urgent the task and the more congested the node are, the smaller the search step size is, so that the search focuses on the vicinity of low-latency and high-resource nodes, and define the fruit fly individual p k The search step size in each round of search iteration is as follows: Among them, λ j k is the search step size of the Drosophila individual p k when processing task t j , representing the local search range of the Drosophila individual p k in the offloading node space. λ0 is the initially set search reference step size, defining the average jump amplitude of Drosophila at the initial stage of task offloading optimization. τ j is the real-time requirement of task t j , representing the maximum allowable response delay of the task. τ max is the maximum allowable delay threshold set by the system. θ is the node load perception adjustment coefficient, controlling the coupling strength between the task urgency and the node congestion state. L′ i(k,j) is the normalized task load of the edge computing node i(k,j), representing the relative saturation of the node's current processing tasks. R′ i(k,j) is the normalized remaining computing resource of the edge computing node i(k,j), representing the relative computing ability of the node available for new task processing; S33. Construct a task-aware multi-objective comprehensive performance function. The task-aware multi-objective comprehensive performance function strongly couples task features with node states, reflects the combined effects of transmission bottlenecks, computing power matching, and load pressure during the task offloading process, and is applied to fruit fly individual p k For task t j evaluate the offloading scheme: Among them, F k,j represents the performance cost value of the selected node of the Drosophila individual p k for task t j The smaller the value, the better the solution. R′ i(k,j) , L′ i(k,j) , B′ i(k,j) , D′ i(k,j) are respectively the normalized remaining computing resources, current task load, network bandwidth, and communication delay of the selected node of the Drosophila individual p k for task t j ; w1, w2, w3, w4 are multi-objective weight factors set by the system. S34. Sort the performance cost values F of all individuals k,j and select, for each task t j the edge computing node number with the minimum cost among all individuals S35. Integrate the optimal edge computing node numbers j corresponding to all tasks t to construct a preliminary set C of candidate nodes for task offloading.

5. A task offloading method for edge computing nodes based on fruit fly optimization and ant colony cooperation according to claim 4, characterized in that The S4 includes the following steps: S41. Use each node in the preliminary task offloading candidate node set C to form the node set V in the task offloading connectivity graph structure, and construct the task offloading connectivity graph G=(V, E), where E is the set of communication paths with direct offloading possibility between candidate nodes, reflecting the network connectivity between nodes; S42. For any two candidate nodes n i , n j in the task offloading connectivity graph G, a dynamic offloading cost function is introduced to quantify the comprehensive performance of the task offloading path between nodes: Among them, D i,j represents the average communication delay from node n i to node n j , B i,j represents the available transmission bandwidth between two nodes, L' j is the normalized current task load of node n j , R' j is the normalized remaining computing resource of node n j , Δ i,j represents the real-time cost fluctuation factor caused by network fluctuations and dynamic changes in node states, and μ1, μ2, μ3, μ4 are the weight coefficients preset by the system; S43. Initialize the pheromone matrix τ = {τ i,j |n i , n j ∈V}, and define the heuristic factor matrix η according to the dynamic offloading cost function. The η in the heuristic factor matrix i,j directly reflects the attractiveness of the offloading path from node n i to n j relative to the cost function; S44. Define the dynamic adaptation factor ζ i,j To enhance the selection probability of resource-advantage nodes and adapt to the heterogeneity of each node in the edge computing environment, the dynamic adaptation factor makes it j when the resources of node n are abundant and the load is low, the probability of being selected as an offloading target is significantly increased, thereby alleviating the local resource tension: Among them, ξ is the resource and load sensitive adjustment coefficient; S45. Introduce each artificial ant into the task offloading connectivity graph G to construct the task offloading path. The k-th ant starts from the offloading initiating node and, in the set of candidate nodes that have not been visited , calculate the dynamic transition probability from node n i to node n j . The dynamic transition probability comprehensively reflects the historical goodness of the offloading path, the current expected cost, and the dynamic advantages of node resources, thereby promoting the global and local collaborative optimization of the ant colony algorithm in edge computing task offloading: Among them, α controls the influence degree of pheromone, β controls the role of heuristic factor, and γ controls the importance of the dynamic adaptation factor ζ i,j ; S46. When each ant completes the construction of an offloading path, calculate the total offloading cost: where Ψ k is the cumulative cost of the path constructed by the k-th ant; S47. Update the pheromone matrix according to the total offloading cost of the paths constructed by each ant; where ρ is the pheromone evaporation coefficient, N a is the total number of ants participating in path construction, and Q is the pheromone release intensity constant; S48. Repeat steps S44 to S47 within the preset number of iterations until the convergence condition is reached; Finally, select the path with the lowest score from all the constructed offloading paths, and the terminal node of the path with the lowest score is determined as the target node for the final task offloading.

6. A task offloading method for edge computing nodes based on fruit fly optimization and ant colony collaboration according to claim 5, characterized in that The S5 includes the following steps: S61. Output the offloading scheduling result according to the finally determined task offloading path and the target node. The task offloading path is the path Path with the minimum total cost in the constructed path set * , and the target node is the end node of this path S62. For each task t in the task set T to be unloaded j according to the corresponding offloading path and the target node perform allocation and migration; S63. During the task migration process, a multi-hop forwarding mechanism is adopted to orderly transmit the task data along the path to the target node, and the integrity verification and reception confirmation of the task data are performed at each hop; S64. When the target node successfully receives task t j and then initializes the task execution environment according to the task feature data T j and the state vector of this node to complete the offloading task deployment; S65. Record the migration completion time, startup latency, and execution duration of task t j in real time, and compare them with the task real-time constraint τ j to generate a task offloading evaluation metric; S66. If the offloading fails due to path interruption, node abnormality or insufficient resources, trigger the backoff scheduling mechanism, reselect the sub-optimal offloading path and the target node from the candidate node set, and execute the task rescheduling and migration operation.

Images