A distributed edge computing method for data communication

Abstract

The application belongs to the technical field of edge computing, and particularly relates to a distributed edge computing method for data communication, which comprises the following steps: obtaining task information of a terminal device and computing power data of an edge node to create a multi-objective optimization model and a mathematical model framework thereof, obtaining an optimal solution satisfying an optimization target and a constraint condition, constructing a distributed computing network architecture, decomposing a computing task into N heterogeneous subtasks and performing subtask granularity optimization, setting priorities for the optimized subtasks and pre-allocating initial edge nodes; based on a node credit evaluation mechanism of a block chain, dynamically adjusting task allocation weights, realizing dynamic migration of the subtasks among the edge nodes through local edge node decision and central node arbitration, and sending preprocessed data requiring deep processing among the edge nodes to the central node for deep processing and then returning to the edge nodes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of server heat dissipation, and particularly relates to a distributed edge computing method for data communication. BACKGROUND

[0002] In the current digital era, data communication volume is growing explosively, and the widespread application of emerging technologies such as the Internet of Things, artificial intelligence, and big data has led to a sharp increase in the amount of data generated by terminal devices. The traditional cloud computing model transmits all data to remote data centers for processing, which faces problems such as high network latency, high bandwidth pressure, high data transmission cost, and privacy and security risks, and is difficult to meet the needs of application scenarios with high real-time requirements and computing-intensive requirements, such as autonomous driving, industrial automation control, and real-time video processing.

[0003] Edge computing, as a new computing paradigm, deploys computing resources on the network edge close to the data source, and sinks part of the data processing tasks from the cloud to the edge nodes, effectively reducing data transmission delay, relieving network bandwidth pressure, and improving data processing privacy and security. However, existing edge computing technologies still have many deficiencies in task allocation, resource utilization, and network architecture construction. For example, in task allocation, there is a lack of efficient multi-objective optimization mechanism, making it difficult to simultaneously consider multiple key indicators such as task processing delay, resource utilization rate, and energy consumption; in resource utilization, it is unable to reasonably schedule and allocate resources according to the dynamic state of terminal devices and edge nodes, resulting in frequent occurrences of resource waste or resource shortage; in network architecture construction, the traditional edge computing network architecture has poor flexibility and adaptability, making it difficult to cope with complex and variable network environments and diverse application requirements. In addition, when edge nodes fail or go offline, existing technologies cannot well realize the continuity of computing relay and data processing, affecting the stability and reliability of the entire system.

[0004] With the development of distributed computing technology, distributed edge computing has gradually become a research hotspot, but how to build an efficient, reliable, and secure distributed edge computing network to achieve reasonable allocation and dynamic scheduling of computing tasks remains a key problem to be solved. SUMMARY

[0005] The purpose of the present application is to provide a distributed edge computing method for data communication to build an efficient, reliable, and secure distributed edge computing network and achieve reasonable allocation and dynamic scheduling of computing tasks.

[0006] To solve the above technical problems, the technical solution adopted by the present application is as follows:

[0007] A distributed edge computing method for data communication, comprising the following steps:

[0008] S1: collecting task information of each terminal device and computing power data of edge nodes in real time, creating a multi-objective optimization model based on the task information of each terminal device and the computing power data of edge nodes;

[0009] S2: setting optimization objectives and constraint conditions of the multi-objective optimization model, and constructing a mathematical model framework of the multi-objective optimization model, solving the multi-objective optimization model by using a specified optimization algorithm, and obtaining an optimal solution satisfying the optimization objectives and constraint conditions;

[0010] S3: constructing a distributed computing network architecture in real time based on the optimal solution, determining the roles and task allocation relationships of each edge node in the network, and constructing a distributed edge computing network composed of multiple edge nodes and a central node;

[0011] S4: decomposing the computing task into N heterogeneous sub-tasks according to the terminal device computing power state and network delay, and performing sub-task granularity optimization, setting priorities for the optimized sub-tasks and pre-allocating initial edge nodes;

[0012] S5: based on a node credit evaluation mechanism of a blockchain, dynamically adjusting task allocation weights, and simultaneously realizing dynamic migration of sub-tasks between edge nodes through local edge node decision-making and central node arbitration;

[0013] S6: when an edge node is offline, a preset diffusion scheme is used to implement preprocessing of all sub-tasks, and each edge node sends preprocessed data that needs to be deeply processed to the central node for deep processing and then returns the data to the edge node;

[0014] S7: the edge node returns the preprocessed data that does not need to be deeply processed directly to the terminal device, and returns the deeply processed data to the terminal device.

[0015] Preferably, the specific process of setting optimization objectives and constraint conditions of the multi-objective optimization model and constructing a mathematical model framework of the multi-objective optimization model in step S2 is as follows:

[0016] S21: setting optimization objectives of the multi-objective optimization model, including minimizing task processing delay, maximizing resource utilization, and minimizing energy consumption;

[0017] S22: setting constraint conditions of the multi-objective optimization model, including computing resource upper limit constraint, network bandwidth limitation constraint, and task data security transmission requirement constraint;

[0018] S23: constructing a mathematical model framework of the multi-objective optimization model based on the optimization objectives and constraint conditions, defining decision variables of the mathematical model framework, and constructing objective functions and constraint conditions thereof.

[0019] Preferably, the detailed execution process of step S23 is as follows:

[0020] S231: define the decision variables of the mathematical model framework of the multi-objective optimization model as x ij When the task i ( i =1, 2, ⋯, n ) is assigned to the edge node j ( j =1, 2, ⋯, m ), x ij =1; otherwise x ij =0, let y ij be a continuous variable, representing the amount of computing resources occupied by the task i when assigned to the edge node j ;

[0021] S232: build the objective function of the mathematical model framework of the multi-objective optimization model:

[0022] ;

[0023] ;

[0024] ;

[0025] wherein Z1 is the objective function of minimizing the task processing delay, Z2 is the objective function of maximizing the resource utilization rate, Z3 is the objective function of minimizing the energy consumption, is the queuing waiting time of the task i at the edge node j , is the time of transmitting the task i from the terminal device to the edge node j , is the actual computing processing time of the task i at the edge node j , α j , β j , gamma j are the weight coefficients of CPU, storage, and network resource utilization rates, respectively; R j c , R j s , R j b are the edge nodej The total amount of CPU, storage, and network resources; s ij 、 b ij Respectively for tasks i The amount of storage resources and network bandwidth occupied on edge node j, e j ( y ij ) is an edge node j Processing tasks i Based on the amount of computing resources occupied y ij Energy consumption function;

[0026] S233: Constraints for establishing the mathematical model framework of the target optimization model:

[0027] Computing resource upper limit constraints:

[0028] ;

[0029] Network bandwidth limitations and constraints:

[0030] ;

[0031] Task data security transmission requirements and constraints:

[0032] ;

[0033] Among them, Security i , j ) indicates a task i At the edge node j The safety standard function that is satisfied during processing returns 1 if satisfied, otherwise returns 0;

[0034] Task assignment uniqueness constraint:

[0035] .

[0036] Preferably, in step S2, the specific process of using a specified optimization algorithm to solve the multi-objective optimization model and obtain the optimal solution that meets the optimization objectives and constraints is as follows:

[0037] S24: Use matrix coding to represent each task as [ x 1, x 2,..., x n ] T ,in x i Corresponding edge node iThe dynamic gene length mechanism is introduced to adapt to the topology change of variable node quantity;

[0038] S25: An initial solution set is generated based on the node capability map to ensure that:

[0039] , and 0≤ x j ≤ C i max ;

[0040] C i max The maximum bearing coefficient of the node i ;

[0041] S26: An ε-dominance screening method is used to retain high-quality solutions with a distance greater than a threshold value ε, establish a solution set archive, record the evolution trajectory of the historical optimal solution, calculate the 95% confidence interval of the target value through Bootstrap sampling, and set a disturbance key parameter to verify the robustness of the optimal solution.

[0042] Preferably, in step S3, the specific process of constructing a distributed computing network architecture based on the optimal solution in real time and dynamically is as follows:

[0043] S31: A center node with global scheduling capability is deployed based on node performance, a dynamic game algorithm framework is loaded, and a resource database and a topology map are initialized;

[0044] S32: The edge node reports the metadata of computing power, storage, and location to the center node through a lightweight protocol to form an initial resource pool;

[0045] S33: Based on real-time load and task demand, the center node calculates the optimal role of each node, including data preprocessing nodes, model inference nodes, and relay nodes, through a potential game algorithm, and periodically evaluates the node state, sets the node role redistribution trigger condition, and triggers role redistribution when the trigger condition is met.

[0046] Preferably, in step S4, the computing task is decomposed into N heterogeneous subtasks according to the terminal device computing power state and network delay, and subtask granularity optimization is performed, and the specific process of setting priority and pre-allocating initial edge nodes for the optimized subtasks is as follows:

[0047] S41: The CPU / GPU computing power, memory, and network delay data of the node are collected through a performance probe to construct a resource feature matrix;

[0048] S42: Based on the resource feature matrix, the genetic algorithm is used to optimize the subtask granularity according to the task type, to ensure that the computing task granularity matches the node computing power;

[0049] S43: Based on the task dependency graph, the key subtasks are identified, and the highest priority is given to reduce the overall time delay. The priority weight is dynamically adjusted according to the business SLA requirements. According to the resource requirements of the subtasks, the Hungarian algorithm is used to match the remaining computing power of the edge nodes with the network topology, and low-delay nodes are preferentially selected. The standby nodes are pre-assigned for high-priority subtasks, and fault-level switching is realized through heartbeat detection.

[0050] Preferably, the specific process of step S42 based on the resource feature matrix and the genetic algorithm to optimize the subtask granularity according to the task type is as follows:

[0051] S421: Define the subtask granularity coding range according to the computing-intensive or IO-intensive;

[0052] S422: Encode the subtask granularity as a chromosome, use real number coding to adapt to continuous parameter optimization, take task processing time delay, resource utilization rate and energy consumption as optimization objectives, dynamically adjust weight coefficients, and iterate optimization convergence to optimal solution through arithmetic crossover and elite reservation strategy;

[0053] S423: According to the optimized subtask demand, match the most suitable node in the resource matrix.

[0054] Preferably, the specific process of step S5 based on the node credit evaluation mechanism of the blockchain to dynamically adjust the task allocation weight, and to realize the dynamic migration of subtasks between edge nodes through local edge node decision and central node arbitration is as follows:

[0055] S51: Record the historical task completion rate, response delay, and computing power stability data of the node through the blockchain, and construct a dynamic credit scoring model;

[0056] S52: Use fuzzy logic algorithm to map credit score to task allocation weight, high-credit nodes are preferentially assigned key subtasks;

[0057] S53: Based on real-time load and credit threshold, local nodes trigger subtask migration request, verify migration rationality through smart contract, prevent malicious node manipulation, and chain arbitration results for storage;

[0058] S54: Based on the Hungarian algorithm, match the computing power of the target node with the network state for the migrated task, and synchronize the task context through lightweight blockchain during the migration process to ensure consistency.

[0059] Preferably, the specific process of step S6 is as follows:

[0060] S61: Detect the node state every specified time through the 5G MEC network, determine as offline if the timeout specified number of times, diffuse the sub-tasks of the offline node to the adjacent specified edge node according to the preset topology according to the priority, and ensure the preprocessing continuity;

[0061] S62: The online node performs data preprocessing, compresses the data volume to a specified proportion of the original data, labels the complex tasks that need central processing through metadata, and attaches the node computing power requirement description;

[0062] S63: Adopt low-delay optical fiber channel transmission to mark data, the center GPU cluster completes AI inference / massive calculation, and the processing result is distributed to the original edge node after being verified by the edge gateway.

[0063] The beneficial effects of the application include:

[0064] The distributed edge computing method for data communication provided by the application can accurately and reasonably allocate and schedule tasks by creating a multi-objective optimization model through real-time collection of terminal device task information and edge node computing power data, setting optimization objectives such as minimizing task processing delay, maximizing resource utilization, and minimizing energy consumption, and combining a specified optimization algorithm to solve the optimal solution. Compared with the traditional method, the task processing delay is greatly reduced, the task execution efficiency is improved, unnecessary energy consumption is reduced, and the overall resource utilization is improved. In actual application scenarios, for tasks such as real-time automatic driving and industrial automation control, the data can be processed in time to ensure the stable operation and safety of the system. In network architecture construction and management, the distributed computing network architecture is dynamically constructed based on the optimal solution in real time, and the role and task allocation relationship of each edge node are determined. The center node uses a dynamic game algorithm framework, combines the metadata reported by the edge node to form a resource pool, and calculates the optimal role of the node through a potential game algorithm, so that the network architecture can be flexibly adjusted according to the real-time load and task demand, significantly enhancing the adaptability and flexibility of the network. It can effectively cope with complex and variable network environments and diversified application requirements, better meet the computing needs in different scenarios, and improve the overall performance and reliability of the network compared with the traditional fixed architecture edge computing network.

[0065] In the sub-task processing and allocation link, the computing task is decomposed and the sub-task granularity is optimized according to the terminal device computing power state and network delay, the genetic algorithm is used to optimize the sub-task granularity, the sub-task granularity is matched with the node computing power, and the priority and pre-allocated node processing mode are set according to the task dependency graph and business SLA requirements, so that the computing task can be efficiently executed on the most suitable node, the mutual interference between tasks is reduced, and the efficiency and quality of task processing are further improved, especially suitable for complex computing-intensive or IO-intensive task scenarios.

[0066] In terms of task dynamic management and node failure response, the node credit evaluation mechanism based on the blockchain dynamically adjusts the task allocation weight, realizes the dynamic migration of sub-tasks through local edge node decision and center node arbitration, guarantees the fairness and rationality of task allocation, improves the security and stability of the system, and prevents malicious node manipulation. When the edge node is offline, the preset diffusion scheme realizes the relay of calculation and data preprocessing, the center node deeply processes specific data, ensures that all sub-tasks can be completed smoothly, maintains the continuity of data processing, and greatly improves the fault tolerance and reliability of the system when facing node failure compared with the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 A flowchart of the distributed edge computing method for data communication of the present application.

[0068] Figure 2 A flowchart of the mathematical model framework for constructing a multi-objective optimization model of the present application.

[0069] Figure 3 A distributed computing network architecture diagram of the present application. DETAILED DESCRIPTION

[0070] The following will be described in conjunction with the accompanying Figures 1-3 Further detailed description of the present application:

[0071] Example 1

[0072] Referring to the accompanying Figure 1 A distributed edge computing method for data communication, comprising the following steps:

[0073] S1: Real-time collection of task information of each terminal device and computing power data of edge nodes, creation of a multi-objective optimization model based on the task information of each terminal device and the computing power data of edge nodes. The distributed computing network architecture of the present application is not fixed, but is dynamically constructed based on real-time task information of each terminal device and computing power and other related data of edge nodes. That is, the distributed computing network architecture of the present application is dynamically constructed according to the real-time state of each terminal device and each node, and the basis of dynamic construction is the multi-objective optimization model constructed in step S2.

[0074] S2: Setting the optimization target and constraint condition of the multi-objective optimization model, and constructing the mathematical model framework of the multi-objective optimization model, solving the multi-objective optimization model by using a specified optimization algorithm, and obtaining the optimal solution satisfying the optimization target and constraint condition. The multi-objective optimization model takes the minimum task processing delay, the maximum resource utilization rate and the minimum energy consumption as the optimization target, and the constraint conditions are the upper limit constraint of computing resources, the network bandwidth limitation constraint and the task data safety transmission requirement constraint.

[0075] S3: Real-time dynamic construction of distributed computing network architecture based on optimal solution, determination of the role of each edge node in the network and the task allocation relationship, and construction of a distributed edge computing network composed of multiple edge nodes and a center node. The distributed computing network architecture constructed based on the solution of the multi-objective optimization model can realize low task processing delay, high resource utilization and low energy consumption in the network under the constraints of computing resource upper limit, network bandwidth limit and task data security transmission requirement.

[0076] S4: According to the terminal device computing power state and network delay, the computing task is divided into N heterogeneous sub-tasks and sub-task granularity optimization is performed, the priority of the optimized sub-tasks is set and the initial edge node is pre-assigned. The computing task is divided into N heterogeneous sub-tasks and sub-task granularity optimization is performed, which provides a basis for subsequent priority setting and task allocation, and through the pre-assigned edge node, it is convenient to realize the cooperative processing of large-scale tasks.

[0077] S5: Based on the node credit evaluation mechanism of the blockchain, the task allocation weight is dynamically adjusted, and the dynamic migration of the sub-tasks between the edge nodes is realized through the local edge node decision and the center node arbitration.

[0078] S6: When the edge node is offline, a preset diffusion scheme is used to realize the preprocessing of all sub-tasks, and the edge nodes send the preprocessing data which needs to be deeply processed to the center node for deep processing and then return to the edge node.

[0079] S7: The edge node returns the preprocessing data which does not need to be deeply processed directly to the terminal device, and returns the deeply processed data to the terminal device.

[0080] In this embodiment, the terminal device task information is collected by embedded sensors and lightweight monitoring agents, real-time collection of task types including computing-intensive and IO-intensive, data size, real-time requirement, security level and other multi-dimensional features, using time window sliding sampling mechanism, combining Kalman filtering algorithm to predict task growth trend, generating task feature vector. The edge node computing power data is collected by deploying a distributed resource monitoring system, periodically collecting CPU / GPU utilization, memory usage, network bandwidth occupation, remaining storage capacity and other dynamic indicators, using DBSCAN density clustering algorithm to identify computing power fluctuation abnormal points, and constructing node resource state time series model.

[0081] The node history task completion rate, response delay, computing power stability and other data are recorded by the blockchain to construct a dynamic credit scoring model. A fuzzy logic algorithm is used to map the credit score to a task allocation weight, so that high-credit nodes can obtain key sub-tasks preferentially. The nodes are allocated reasonably according to their actual performance, so as to ensure that resources are tilted towards more reliable and efficient nodes, improve the overall efficiency of task processing, ensure the fairness of the task allocation process, and encourage each node to actively improve its performance.

[0082] The credit data of the nodes and the arbitration results of the task migration are recorded by the blockchain and stored on the chain for evidence, so as to prevent malicious nodes from tampering with the data or manipulating the task allocation process. In the sub-task migration process, the smart contract verifies the rationality of the migration, and only the migration request that meets the rules can be executed, thereby further ensuring the security of the system and reducing the risk of system failure or data leakage caused by malicious attacks.

[0083] The local edge node triggers a sub-task migration request based on the real-time load and the credit threshold, and the central node arbitrates. Through the combination of local decision and central arbitration, the system can quickly respond according to the real-time state of the nodes. When the load of a node is too high or the performance of the node decreases, the sub-task can be migrated to a more suitable node in time to avoid task backlog and ensure the continuous and efficient execution of the task. At the same time, based on the Hungarian algorithm, the computing power and network state of the target node are matched with the migrated task. In the migration process, the task context is synchronized through a lightweight blockchain to ensure the consistency and continuity of the task execution, thereby improving the ability of the system to cope with complex tasks and dynamic environments. This mechanism enables the system to maintain overall operation through dynamic sub-task migration when some nodes fail or have unstable performance. The tasks of low-credit or offline nodes can be redistributed to other reliable nodes, ensuring that the computing task will not be interrupted due to individual node problems, thereby enhancing the fault tolerance and reliability of the system and ensuring the stable operation of the distributed edge computing network.

[0084] Embodiment 2

[0085] Based on embodiment 1, referring to Figure 2 , the specific process of setting the optimization objectives and constraint conditions of the multi-objective optimization model in step S2 and constructing the mathematical model framework of the multi-objective optimization model is as follows:

[0086] S21: The optimization objectives of the multi-objective optimization model include minimizing the task processing delay, maximizing the resource utilization rate, and minimizing the energy consumption.

[0087] S22: The constraint conditions of the multi-objective optimization model include the upper limit constraint of the computing resources, the network bandwidth limitation constraint, and the task data security transmission requirement constraint.

[0088] S23: Construct a mathematical model framework of the multi-objective optimization model based on the optimization target and the constraint condition, define decision variables of the mathematical model framework, and construct an objective function and a constraint condition thereof.

[0089] The above process of constructing the mathematical model framework of the multi-objective optimization model collectively serves to construct an optimization decision framework capable of balancing multi-dimensional targets and coping with complex constraint conditions, and the core technical purpose is to realize efficient resource configuration and intelligent task scheduling in a distributed edge computing environment.

[0090] In the edge computing environment, the three targets of task processing delay, resource utilization rate and energy consumption are related to each other and in conflict. For example, reducing the delay may require increasing resource allocation, thereby causing the energy consumption to rise, and improving the resource utilization rate may cause the delay of some nodes to increase due to task concentration. The traditional single-target optimization method cannot effectively cope with such complex trade-off relationships.

[0091] The constraint conditions such as the upper limit of computing resources, network bandwidth limitation and data security requirements in the edge computing network have the characteristics of dynamic change. For example, the network bandwidth fluctuates with the change of user access volume, and the security requirements also differ due to different task types. How to make the optimal decision under these dynamic constraint conditions is a key challenge. The edge computing network is usually composed of a large number of heterogeneous nodes, including devices with different computing capabilities, storage capacities and network bandwidths. How to realize global optimization in such a heterogeneous environment and avoid local optimal solution is the key to improving the performance of the entire system.

[0092] The technical solution realized by the present application quantitatively models the three targets of minimizing task processing delay, maximizing resource utilization rate and minimizing energy consumption, and establishes a mathematical model capable of comprehensively considering multiple targets. This multi-objective optimization method can find a set of Pareto optimal solutions, providing multiple selection schemes for decision makers. The constraint conditions such as the upper limit of computing resources, network bandwidth limitation and task data security transmission requirement are converted into mathematical expressions, ensuring that the optimization process is carried out within the feasible solution space.

[0093] Decision variable and objective function design: the decision variable is defined x ij and y ij represent the task allocation relationship and the resource occupation amount, and the corresponding objective function and constraint condition are constructed, so that the multi-objective optimization model can accurately reflect the resource allocation and task scheduling problem in the edge computing environment.

[0094] By considering multiple optimization objectives and constraint conditions at the same time, efficient configuration of computing resources, storage resources and network resources is realized. Under the premise of meeting various constraint conditions, resource utilization is improved as much as possible, energy consumption is reduced, and task processing delay is reduced. The optimized task allocation and resource scheduling strategy can significantly improve the performance of the entire edge computing system. For example, by minimizing the task processing delay, the response speed of real-time applications can be improved; by maximizing the resource utilization, the system construction cost can be reduced; by minimizing the energy consumption, the device service life can be prolonged and energy consumption can be reduced. The multi-objective optimization model can adapt to the dynamic changes of the edge computing environment, including node joining or exiting, load fluctuation, network state change, etc. By adjusting the task allocation and resource scheduling strategy in real time, the system can maintain good performance under various dynamic conditions.

[0095] The detailed execution process of step S23 is as follows:

[0096] S231: define the decision variables of the mathematical model framework of the multi-objective optimization model as x ij When the task i ( i =1,2,⋯, n ) is allocated to the edge node j ( j =1,2,⋯, m ), x ij =1; otherwise x ij =0, let y ij be a continuous variable, representing the amount of computing resources occupied by the task i when allocated to the edge node j ;

[0097] S232: construct the objective function of the mathematical model framework of the multi-objective optimization model:

[0098] ;

[0099] ;

[0100] ;

[0101] wherein Z1 is the objective function of minimizing the task processing delay, Z2 is the objective function of maximizing the resource utilization, and Z3 is the objective function of minimizing the energy consumption, is the queuing waiting time of the task i at the edge node j , is the transmission time of the task i from the terminal device to the edge nodej the time, for the task i on the edge node j the actual computing processing time, α j , β j , gamma j respectively, the weight coefficients of CPU, storage, network resource utilization; R j c , R j s , R j b respectively, the total amount of CPU, storage, network resources of the edge node j s ij , b ij respectively, the storage resource amount and network bandwidth occupied by the task i on the edge node j, e j y ij is the energy consumption function based on the amount of computing resources occupied j when the edge node i processes the task y ij

[0102] S233: Establish the constraint conditions of the mathematical model framework of the target optimization model:

[0103] Upper limit constraint of computing resources:

[0104] ;

[0105] Network bandwidth limit constraint:

[0106] ;

[0107] Task data security transmission requirement constraint:

[0108] ;

[0109] wherein, Security( i , j ) represents the security standard function met when the task i is processed on the edge node j , and returns 1 if met, otherwise returns 0;

[0110] ​​​Task allocation uniqueness constraint:

[0111] .

[0112] The specific process of solving the multi-objective optimization model by using the specified optimization algorithm in step S2 to obtain the optimal solution satisfying the optimization objective and constraint condition is as follows:

[0113] S24: Each task is represented as x 1, x 2,..., x n T , wherein x i corresponding to the task allocation proportion of the edge node i , a dynamic gene length mechanism is introduced to adapt to the topology changes of variable node quantity;

[0114] S25: Generate an initial solution set based on the node capability map to ensure that:

[0115] , and 0≤ x j ≤ C i max ;

[0116] C i max is the maximum bearing coefficient of node i ;

[0117] S26: Use the ε-dominance screening method to retain high-quality solutions with a distance greater than the threshold value ε, establish a solution set archive, record the evolution trajectory of the historical optimal solution, calculate the 95% confidence interval of the target value through Bootstrap sampling, and set the disturbance key parameter to verify the robustness of the optimal solution.

[0118] Example 3

[0119] On the basis of example 1 or example 2, the specific process of constructing a distributed computing network architecture in real time based on the optimal solution in step S3 to determine the role and task allocation relationship of each edge node in the network, and to construct a distributed edge computing network composed of multiple edge nodes and a central node is as follows:

[0120] S31: Deploy a central node with global scheduling capability based on node performance, load a dynamic game algorithm framework, and initialize the resource database and topology map.

[0121] ​By acquiring the computing power of each node, including CPU / GPU frequency, core number, and memory capacity, network bandwidth, stability, a center node candidate set is established, and the node with the highest comprehensive score in the center node candidate set is selected as the center node. Define the strategy space: S = {s1(data preprocessing), s2(model inference), s3(relay transmission)}, set the revenue function, use graph data structure to represent network topology: G=(V, E) edge weight calculation, use Dijkstra algorithm to precompute the shortest path matrix to build the topology map.

[0122] S32: The edge node reports the metadata of computing power, storage, and location to the center node through a lightweight protocol to form an initial resource pool.

[0123] Based on the CoAP protocol extension, support UDP multicast communication, message format is:

[0124] { "node_id": "0xABC123",

[0125] "timestamp": 1693542180,

[0126] "cpu_available": 75%,

[0127] "memory_available": 60%,

[0128] "storage_available": 85GB,

[0129] "location": {"lat": 30.1234, "lng": 120.5678},

[0130] "status": "online"}

[0131] Adopt MsgPack binary serialization format, data compression, sliding time window filter obsolete data, Kalman filter smooth resource fluctuation data, based on GeoHash algorithm to build 2D space index, build multi-dimensional B+ tree index, support range query.

[0132] S33: Based on real-time load and task demand, the center node calculates the optimal role of each node through potential game algorithm, including data preprocessing node, model inference node, relay node, and periodically evaluates node status, sets node role reallocation trigger condition, when the reallocation trigger condition is met, triggers role reallocation.

[0133] A system state vector is constructed, the system vector includes a node load vector and a resource vector of each node, all node roles are initialized, and the node state is periodically evaluated, the node role redistribution trigger condition is set, and the role redistribution is triggered when the redistribution trigger condition is met. The flow of redistribution is: marking the nodes that need to be adjusted, local redistribution: finding a replacement in the neighbor of the failed node, global rebalancing: if local adjustment fails, triggering global role recalculation.

[0134] The created distributed computing network architecture schematic diagram is shown in Figure 3 The center node is determined by real-time acquisition of the computing power of each node. The center node may change according to the real-time state of each node, so the center node is dynamic, and the edge node is also dynamic. The tasks generated by the terminal device are transmitted to the edge node, the edge node performs preliminary processing and data compression, and then transmits them to the center node. The center node processes the tasks through the task scheduling module to generate a task queue, and the center node processes the tasks according to the task queue.

[0135] The specific process of decomposing the computing task into N heterogeneous subtasks and performing subtask granularity optimization according to the terminal device computing power state and network delay in step S4 is as follows:

[0136] S41: Collect node CPU / GPU computing power, memory and network delay data through performance probes, and construct a resource feature matrix;

[0137] S42: Based on the resource feature matrix, the subtask granularity is optimized according to the task type by using a genetic algorithm to ensure that the computing task granularity matches the node computing power;

[0138] S43: Based on the task dependency graph, identify the key subtasks, assign the highest priority to reduce the overall delay, dynamically adjust the priority weight according to the business SLA requirement, match the remaining computing power of the edge node with the network topology according to the subtask resource requirement by using the Hungarian algorithm, preferentially select low-delay nodes, pre-allocate standby nodes for high-priority subtasks, and realize fault second-level switching through heartbeat detection.

[0139] The specific process of optimizing the subtask granularity according to the task type by using a genetic algorithm based on the resource feature matrix in step S42 is as follows:

[0140] S421: Define the subtask granularity coding range according to the computing-intensive or IO-intensive.

[0141] S422: Encode the subtask granularity into a chromosome, adopt real number coding to adapt to continuous parameter optimization, take the task processing delay, resource utilization, and energy consumption as the optimization objectives, dynamically adjust the weight coefficient, and converge to the optimal solution through arithmetic crossover and elite reservation strategy iteration optimization.

[0142] The process of iterative optimization is as follows:

[0143] Initialize the population, randomly generate N chromosomes, calculate the fitness of the initial population, apply elite reservation, save the optimal K chromosomes, generate a mating pool through tournament selection, perform arithmetic crossover on the chromosomes in the mating pool, perform Gaussian mutation on the crossover results, combine the elite chromosomes and the newly generated chromosomes, calculate the fitness, and if the convergence condition is met, exit the loop, and finally return the optimal chromosome.

[0144] S423: According to the optimized subtask demand, match the most suitable node in the resource matrix.

[0145] Embodiment 4

[0146] On the basis of embodiment 1 or embodiment 2 or embodiment 3, the node credit evaluation mechanism based on the blockchain in step S5 dynamically adjusts the task allocation weight, and the specific process of realizing the dynamic migration of subtasks between edge nodes through local edge node decision and central node arbitration is as follows:

[0147] S51: Record the historical task completion rate, response delay, and computing power stability data of the node through the blockchain, and construct a dynamic credit scoring model;

[0148] S52: Use fuzzy logic algorithm to map credit score to task allocation weight, high credit nodes get critical subtasks first;

[0149] S53: Local nodes trigger subtask migration request based on real-time load and credit threshold, verify migration rationality through smart contract, prevent malicious node manipulation, and chain the arbitration result for storage;

[0150] S54: Based on the Hungarian algorithm, match the computing power of the target node and the network state of the migrated task, and synchronize the task context through lightweight blockchain during the migration process to ensure consistency.

[0151] Data collection modules are deployed at each edge node to monitor the task completion rate, response delay, and computing power stability of the node in real time. The task completion rate is obtained by calculating the ratio of the number of successfully processed tasks to the total number of received tasks. The response delay records the time interval from task assignment to node result return. The computing power stability is determined by the fluctuation range of CPU and GPU usage during task processing. The collected data is encrypted by hashing and broadcast to other nodes through the P2P network of the blockchain. After verifying the data validity, each node packages it into a new block to achieve an unalterable record of data.

[0152] A dynamic credit scoring model is constructed using the weighted average method, with historical task completion rate C rate , response delay D delay , and computing power stability S stable , and the corresponding weight coefficients are α , β , gamma α β gamma =1. The credit score S core .

[0153] The calculation formula is as follows:

[0154] .

[0155] where C rate ∈[0, 1], D delay , and the reciprocal of the average response delay over a period of time is taken, S stable ∈[0,1]. The weight coefficient is dynamically adjusted according to different application scenarios through the analytic hierarchy process. At the same time, to reflect the importance of the node's recent performance, a time decay factor delta , (0 delta <1) is introduced to weight the historical data. As time passes, the data weight gradually decreases.

[0156] The credit score is divided into multiple fuzzy sets, such as "low credit", "medium credit", and "high credit". The corresponding membership functions are defined, such as the Z-type membership function for "low credit", the trapezoidal membership function for "medium credit", and the S-type membership function for "high credit". According to the credit score of the node, the membership degree of different fuzzy sets is calculated.

[0157] A fuzzy rule base is established:

[0158] ​​​If the node credit is "high credit", the task allocation weight is "high";

[0159] If the node credit is "medium credit", the task allocation weight is "medium";

[0160] If the node credit is "low credit", the task allocation weight is "low".

[0161] The barycentric method is used to convert the fuzzy output into a specific task allocation weight value. By calculating the barycentric coordinates of the fuzzy set, the final task allocation weight is obtained, so that the high credit node obtains a higher weight and thus obtains the key subtask in priority.

[0162] The local edge node monitors its load in real time. When the load exceeds the set threshold and the node credit score is lower than the credit threshold, the subtask migration request is triggered. The migration request contains the relevant information of the subtask to be migrated, such as task type, data volume, and computing demand. The subtask migration request is sent to the blockchain network, and the smart contract verifies the migration request according to the preset rules. The rules include whether the load of the target node is lower than a certain threshold, whether the credit score of the target node meets the requirements, and whether the migration will cause the interruption of the source node task processing, etc. Only the verified migration request will be executed to prevent malicious nodes from interfering with the system operation through false migration requests. If there is a dispute in the migration request, such as multiple nodes competing for the same subtask, the central node will arbitrate. The central node makes an arbitration decision based on the credit score, load condition, and task priority of each node, and the arbitration result is recorded on the blockchain through the smart contract to ensure the transparency and traceability of the entire migration process.

[0163] The computing power and network state matrix of the target node is constructed, and the matrix elements include the CPU remaining computing power, memory remaining capacity, network bandwidth remaining amount, etc. The computing demand and network transmission demand of the subtask to be migrated are taken as the cost function, and the Hungarian algorithm is used to solve the minimum cost allocation scheme to determine the optimal matching relationship between the subtask and the target node. In the task migration process, the lightweight blockchain technology is used to synchronize the task context information, including the execution progress, intermediate calculation result, and data dependency relationship of the subtask. The task context data is linked to the blockchain through the hash pointer, and the target node can obtain the complete and accurate task context by verifying the hash value, ensuring that the task can continue to execute seamlessly after migration and maintaining the consistency of task processing.

[0164] The specific process of step S6 is as follows:

[0165] S61: Detect the node state through the 5G MEC network every specified time, and determine as offline if it is timed out for a specified number of times. According to the preset topology, the subtasks of the offline node are diffused to the adjacent specified edge nodes according to the priority to ensure the continuity of preprocessing.

[0166] S62: The online node performs data preprocessing, compresses the data volume to a specified proportion of the original data, and labels the complex tasks that need to be processed by the center through metadata, and attaches a node computing power requirement specification;

[0167] S63: Adopt low-delay fiber channel transmission to mark data, and the center GPU cluster completes AI inference / massive calculation, and the processing results are distributed to the original edge node after verification by the edge gateway.

[0168] In the 5G MEC (Mobile Edge Computing) network, a dedicated heartbeat detection module is deployed. This module sends a state detection request to each edge node every T time, and the node returns a response packet containing its CPU usage, memory occupation, network connection state, etc. If the center node does not receive a response from a certain node within the specified T time, it is considered a timeout; when the same node is continuously timed out N times, it is determined that the node is offline.

[0169] A sub-task priority evaluation system is established, and different priorities are given according to the real-time requirements of the task, the importance of the business, the data processing stage, etc. For example, real-time data processing tasks for emergency obstacle avoidance in autonomous driving scenarios are set to the highest priority; while log analysis tasks are set to a lower priority. Priority is identified by a number, and the larger the number, the higher the priority.

[0170] According to the preset network topology, the center node constructs a node adjacency graph. When a node is offline, the center node traverses the sub-task list of the node from high to low, and spreads the sub-task to adjacent edge nodes with a load lower than the threshold value. During the spreading process, the sub-task description information is quickly transmitted through the 5G low-delay communication channel to ensure the continuity of the preprocessing work.

[0171] After the online node receives the data processing task, it first judges the data type. For multimedia data such as images and videos, efficient compression algorithms are used to compress the data volume to R proportion of the original data; for structured data, dictionary compression, difference encoding, etc. are used for compression. Establish a metadata labeling specification, and the online node identifies complex sub-tasks that need to be processed by the center node through AI inference or massive calculation by analyzing the task processing logic. For these tasks, special identifiers are added to the metadata, and the computing resources required by the task are specified in detail. For example, for an image recognition task, it is labeled that it needs to use the ResNet-50 model, and at least occupies 2 GPU cores and 8GB video memory.

[0172] A dedicated low-latency optical fiber network is constructed to connect each edge node and the center node. After the center node receives the preprocessed data marked, it distributes the data to the corresponding GPU cluster computing nodes according to the task requirements in the metadata. In the transmission process, a high-speed data transmission protocol is used to reduce data transmission delay and network congestion. The GPU cluster dynamically allocates computing resources according to the task type and computing power requirements. For AI inference tasks, the corresponding deep learning model is loaded to process the data; for large-scale computing tasks, a distributed computing framework is used for parallel computing. After processing, the results are packaged, including processing result data, task execution status and other information.

[0173] After the edge gateway receives the processing results returned by the center node, it first performs integrity and accuracy verification. By comparing the hash value of the calculation result data with the hash value before transmission, it verifies that the data has not been tampered with during transmission; at the same time, according to the task type, the corresponding result validity verification is performed. After verification, the edge gateway distributes the results to the original edge node, and the original edge node continues the subsequent data processing or returns the results to the terminal device according to the results.

[0174] In summary, the distributed edge computing method for data communication provided by the application creates a multi-objective optimization model by collecting terminal device task information and edge node computing power data in real time, sets optimization objectives such as minimizing task processing delay, maximizing resource utilization, minimizing energy consumption, and multiple constraint conditions, and combines a specified optimization algorithm to solve the optimal solution, which can accurately allocate and schedule tasks. Reducing the task processing delay, improving the task execution efficiency, reducing the unnecessary energy consumption, and improving the overall resource utilization. Based on the optimal solution, a distributed computing network architecture is dynamically constructed in real time to determine the role of each edge node and the task allocation relationship. The center node uses a dynamic game algorithm framework to form a resource pool combined with the metadata reported by the edge node, and calculates the optimal role of the node through the potential game algorithm, so that the network architecture can be flexibly adjusted according to the real-time load and task requirements, significantly enhancing the adaptability and flexibility of the network.

[0175] According to the terminal device computing power state and network delay, the task is decomposed and sub-task granularity is optimized. Genetic algorithm is used to optimize the sub-task granularity, so that it matches the node computing power. Combined with task dependency graph and business SLA requirement, priority and pre-distribution node processing mode are set to ensure that the computing task can be efficiently executed on the most suitable node, reducing the mutual interference between tasks. Based on the node credit evaluation mechanism of the blockchain, the task allocation weight is dynamically adjusted, the sub-task dynamic migration is realized through local edge node decision and center node arbitration, which not only guarantees the fairness and rationality of task allocation, but also improves the security and stability of the system. When the edge node is offline, the preset diffusion scheme realizes the computing relay and data preprocessing, and the center node deeply processes the specific data, ensuring that all sub-tasks can be completed smoothly and maintaining the continuity of data processing.

Claims

1. A method for distributed edge computing for data communication, characterized in that, The method comprises the following steps: S1: Real-time collection of task information of each terminal device and computing power data of edge nodes, creation of a multi-objective optimization model based on the task information of each terminal device and the computing power data of edge nodes; S2: Setting optimization objectives and constraint conditions of the multi-objective optimization model, and constructing a mathematical model framework of the multi-objective optimization model, solving the multi-objective optimization model by using a specified optimization algorithm, and obtaining an optimal solution satisfying the optimization objectives and constraint conditions; S3: Real-time dynamic construction of a distributed computing network architecture based on the optimal solution, determination of the roles and task allocation relationships of each edge node in the network, and construction of a distributed edge computing network composed of multiple edge nodes and a central node; S4: According to the terminal device computing power state and network delay, the computing task is decomposed into N heterogeneous subtasks and subtask granularity optimization is performed, the priority of the optimized subtasks is set and the initial edge node is pre-assigned; S5: Based on the node credit evaluation mechanism of the blockchain, the task allocation weight is dynamically adjusted, and the dynamic migration of the subtasks between the edge nodes is realized through local edge node decision and central node arbitration; S6: When the edge node is offline, a preset diffusion scheme is used for computing relay to realize preprocessing of all subtasks, and the edge node sends the preprocessing data requiring deep processing to the central node for deep processing and then returns the data to the edge node; S7: The edge node returns the preprocessing data not requiring deep processing to the terminal device, and returns the deep-processed data to the terminal device; The specific process of step S3 is as follows: S31: A central node with global scheduling capability is deployed based on node performance, a dynamic game algorithm framework is loaded, and a resource database and a topology map are initialized; S32: The edge node reports the metadata of computing power, storage, and location to the central node through a lightweight protocol to form an initial resource pool; S33: Based on real-time load and task demand, the central node calculates the optimal roles of each node, including data preprocessing nodes, model inference nodes, and relay nodes, and periodically evaluates the node state, sets the node role reallocation trigger condition, and triggers the role reallocation when the reallocation trigger condition is met; The specific process of step S6 is as follows: S61: The node state is detected every specified time through the 5G MEC network, and the node is determined to be offline when the timeout specified number of times is reached, the subtasks of the offline node are diffused to adjacent specified edge nodes according to the preset topology, and the preprocessing continuity is ensured; S62: The online node performs data preprocessing, compresses the data volume to a specified proportion of the original data, marks the complex tasks requiring central processing through metadata, and attaches the node computing power demand description; S63: The marked data is transmitted through a low-delay optical fiber channel, the central GPU cluster completes AI inference / massive computing, and the processing result is distributed to the original edge node after verification by the edge gateway. 2.The distributed edge computing method for data communication of claim 1, wherein, The specific process of setting optimization objectives and constraint conditions of the multi-objective optimization model and constructing a mathematical model framework of the multi-objective optimization model in step S2 is as follows: S21: setting the optimization target of the multi-objective optimization model includes minimizing task processing delay, maximizing resource utilization, and minimizing energy consumption; S22: setting the constraint conditions of the multi-objective optimization model includes computing resource upper limit constraint, network bandwidth limit constraint, and task data secure transmission requirement constraint; S23: constructing a mathematical model framework of the multi-objective optimization model based on the optimization target and the constraint conditions, defining the decision variables of the mathematical model framework, and constructing the objective function and its constraint conditions.

3. The distributed edge computing method for data communication according to claim 2, wherein, The detailed execution process of step S23 is as follows: S231: The decision variables of the mathematical model framework of the multi-objective optimization model are defined as x ij , when the task i ( i =1,2,⋯, n ) assigned to edge nodes j ( j =1,2,⋯, m )hour, x ij =1; otherwise x ij =0, let y ij is a continuous variable, indicating the task i Assign to edge nodes j The amount of computing resources used when S232: constructing the objective function of the mathematical model framework of the multi-objective optimization model: ； ； ； Z1 is a target function for minimizing the task processing delay, Z2 is a target function for maximizing the resource utilization, Z3 is a target function for minimizing the energy consumption, is the task i waiting time in the edge node j queue, is the task i transmission time from the terminal device to the edge node j , is the actual computing processing time of the task i on the edge node j , α j , β j , S233: establishing the constraint conditions of the mathematical model framework of the objective optimization model: j are the weight coefficients of CPU, storage, and network resource utilization, respectively; R j c , R j s , R j b are the total amount of CPU, storage, and network resources of the edge node j , s ij , b ij are the storage resource amount and network bandwidth occupied by the task i on the edge node j , e j ( y ij ) is the energy consumption function of the edge node j processing the task i based on the occupied computing resource amount y ij . computing resource upper limit constraint: network bandwidth limit constraint: ； task data secure transmission requirement constraint: ； task allocation uniqueness constraint: ； where Security( i , j ) denotes a security standard function that is satisfied by the task i at the edge node j processing, returning 1 if satisfied, otherwise 0; The specific process of solving the multi-objective optimization model by using the specified optimization algorithm in step S2 to obtain the optimal solution that satisfies the optimization target and the constraint conditions is as follows: 。 4. The distributed edge computing method for data communication according to claim 3, wherein, S25: generating an initial solution set based on the node capability graph to ensure that: S24: each task is expressed as a matrix coding mode [ x 1, x 2,..., x n ] T , wherein x i the task allocation ratio of the corresponding edge node i , a dynamic gene length mechanism is introduced to adapt to the topology change of variable node number; S26: using the ε-dominance screening method to retain high-quality solutions with a distance greater than a threshold value ε, establishing a solution set archive, recording the evolution trajectory of the historical optimal solution, calculating the 95% confidence interval of the target value through Bootstrap sampling, and setting a disturbance key parameter to verify the robustness of the optimal solution. , and 0 ≤ x j ≤ C i max ; C i max For the node i the maximum bearing coefficient; The specific process of decomposing the computing task into N heterogeneous sub-tasks and performing sub-task granularity optimization according to the terminal device computing power state and network delay in step S4 is as follows:

5. The distributed edge computing method for data communication according to claim 1, wherein, S41: collecting node CPU / GPU computing power, memory, and network delay data through performance probes to construct a resource feature matrix; S42: based on the resource feature matrix, using genetic algorithm to optimize sub-task granularity according to task type to ensure that the computing task granularity matches the node computing power; S43: based on the task dependency graph, identifying key sub-tasks and giving them the highest priority to reduce overall delay, dynamically adjusting priority weights according to business SLA requirements, matching edge node remaining computing power and network topology using the Hungarian algorithm according to sub-task resource requirements, preferentially selecting low-delay nodes, pre-allocating standby nodes for high-priority sub-tasks, and achieving fault-second-level switching through heartbeat detection. The specific process of using genetic algorithm to optimize sub-task granularity based on the resource feature matrix according to task type in step S42 is as follows:

6. The distributed edge computing method for data communication according to claim 5, wherein, S421: defining sub-task granularity coding range according to computing-intensive or IO-intensive; S422: coding sub-task granularity as a chromosome, using real number coding to adapt to continuous parameter optimization, taking task processing delay, resource utilization, and energy consumption as optimization targets, dynamically adjusting weight coefficients, and iteratively optimizing and converging to the optimal solution through arithmetic crossover and elite reservation strategy; S423: matching the most suitable node in the resource matrix according to the optimized sub-task requirements. ​ 7. The distributed edge computing method for data communication according to claim 1, wherein, The specific process of the dynamic migration of sub-tasks between edge nodes through local edge node decision-making and central node arbitration in step S5 based on the blockchain-based node credit evaluation mechanism is as follows: S51: Record the historical task completion rate, response delay, and computing power stability data of the node through the blockchain to construct a dynamic credit scoring model; S52: Use a fuzzy logic algorithm to map the credit score to the task allocation weight, and high-credit nodes are given priority in obtaining key sub-tasks; S53: Local nodes trigger sub-task migration requests based on real-time load and credit thresholds, verify the rationality of migration through smart contracts to prevent malicious node manipulation, and arbitrate the results on-chain for evidence; S54: Based on the Hungarian algorithm, match the computing power and network status of the target node for the migrated task, and synchronize the task context during the migration process through a lightweight blockchain to ensure consistency.

Images