An edge cloud computing resource allocation optimization method based on deep learning

Abstract

The application relates to the field of intelligent scheduling and allocation, in particular to an edge cloud computing resource allocation optimization method based on deep learning. The application adopts a time-space prediction algorithm based on multi-head attention and gate mechanism optimization, designs time encoding and space encoding, captures the time sequence characteristics of computing power load and the spatial relationship and interaction between edge server nodes, simultaneously combines the multi-head attention mechanism with the expansion causal convolution, can capture the instantaneous computing power load fluctuation, can mine the long-term trend of the computing power load, and thus predicts the accurate computing power load to provide a reliable basis for subsequent computing power scheduling. The application designs an alternating direction multiplier method based on genetic algorithm optimization, is suitable for nonlinear and multi-constraint optimization problems, can be extended to a larger-scale distributed edge node cloud computing system, simultaneously rapidly approaches the global optimal solution through the genetic algorithm, improves the initial solution quality, accelerates model convergence, and improves the distributed collaborative allocation and scheduling efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent scheduling and allocation, and particularly relates to an edge cloud computing resource allocation optimization method based on deep learning. BACKGROUND

[0002] The world is rapidly entering an intelligent era driven by the Internet of Things and artificial intelligence, and a large number of new applications that are extremely sensitive to delay and computationally intensive have emerged. The traditional centralized cloud computing mode (transferring all data to a remote cloud data center for processing) has been unable to effectively cope with these new demands, and problems such as high network delay and poor allocation of computing resources have become increasingly prominent. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the application provides an edge cloud computing resource allocation optimization method based on deep learning. The application uses a spatio-temporal prediction algorithm based on multi-head attention and gate mechanism optimization to predict and analyze the computing power load demand of terminal equipment, designs time encoding and space encoding, captures the time sequence characteristics of the computing power load and the spatial relationship between the edge server nodes, understands the interaction between the computing power load demands of the edge server nodes, and combines the multi-head attention mechanism with the dilated causal convolution, which can capture the instantaneous computing power load fluctuation and mine the long-term trend of the computing power load, so as to predict accurate computing power load and provide a reliable basis for subsequent computing power scheduling. In view of the problems of high network delay and poor allocation of computing resources in the prior art when processing edge node computing power, the application designs an alternating direction multiplier method based on genetic algorithm optimization. The genetic algorithm provides global search capability to avoid falling into local optimum, and the ADMM alternating direction multiplier method realizes distributed decoupling and reduces the computational complexity. The genetic algorithm is suitable for nonlinear, multi-constraint optimization problems and is compatible with multiple models, and the ADMM allows expansion to larger-scale distributed edge node cloud computing systems. The genetic algorithm quickly approaches the global optimal solution, improves the initial solution quality, speeds up the ADMM convergence, and improves the efficiency of distributed collaborative allocation and scheduling.

[0004] The application provides an edge cloud computing resource allocation optimization method based on deep learning, which specifically comprises the following steps:

[0005] Step S1: edge server deployment, distributed deployment of edge servers, wherein each edge server node can adjust its own computing resources;

[0006] Step S2: communication and data sharing, connecting each edge server through Internet of Things technology to realize information interaction between nodes;

[0007] Step S3: Computing power load prediction, a spatio-temporal prediction algorithm based on multi-head attention and gate mechanism optimization is used to predict the computing power load demand of the terminal device, which specifically includes the following steps:

[0008] Step S31: Data collection, collect the historical time series load demand M of the terminal device to each edge server, where the dimension of M is T*N*1, T represents the step length of the historical time series, N represents the number of edge server nodes, and each node forms a node graph according to the geographical position;

[0009] Step S32: Spatio-temporal encoding, the historical time series load demand M of each edge server node is spatio-temporally encoded to obtain the spatio-temporal encoding, which specifically includes the following steps:

[0010] Step S321: Time encoding, the historical time series is time-encoded:

[0011] ;

[0012] ;

[0013] In the formula, represents the step length of the historical time series, represents the dimension of the hidden layer in the encoding, represents the time of the historical time series, and is the time encoding;

[0014] Step S322: Global structure information collection, the node graph is spatially encoded by using a normalized Laplacian matrix to obtain a global structure matrix;

[0015] Step S323: Spatial encoding, the global structure matrix and the historical time series load demand M are multiplied to obtain the spatial encoding;

[0016] Step S324: Spatio-temporal encoding, the spatial encoding and the time encoding are fused to obtain the spatio-temporal encoding;

[0017] Step S33: Dilated causal convolution processing, the spatio-temporal encoding is dilated causal convolution processed to obtain the time series dependency relationship:

[0018] ;

[0019] In the formula, represents the value of the spatio-temporal encoding at time step t-k, represents the inflation rate, represents the convolution kernel weight, represents the convolution kernel size, which represents the time series dependency relationship;

[0020] Step S34: multi-head attention mechanism processing, multi-head attention processing is performed on the space-time coding and time sequence dependency to obtain a space-time feature, and the following steps are included:

[0021] Step S341: query vector and key vector generation, linear transformation is performed on the space-time coding to generate a query vector and a key vector;

[0022] Step S342: value vector generation, linear coding is performed on the time sequence dependency to generate a value vector;

[0023] Step S343: attention mechanism processing is performed on the query vector, the key vector and the value vector:

[0024] ;

[0025] In the formula, query vector, key vector and value vector, represent the dimensions of the query vector and the key vector, represent an activation function;

[0026] Step S344: multi-head attention optimization, multi-head attention optimization is performed on the result of the attention mechanism processing to obtain a space-time feature;

[0027] Step S35: feature fusion based on a gating mechanism, fusion optimization based on a gating mechanism and residual connection is performed on the space-time feature and the time sequence dependency to obtain an optimized space-time feature:

[0028] ;

[0029] In the formula, represent layer processing, represent layer normalization, represent the space-time feature, represent a Sigmoid activation function, represent the time sequence dependency, represent the space-time coding, represent the optimized space-time feature;

[0030] Step S36: multi-layer perception processing and prediction output, multi-layer perception processing is performed on the optimized space-time feature and an activation prediction is performed to obtain a predicted computing power load demand;

[0031] Step S4: cloud distributed optimization and collaborative control, an alternating direction multiplier method based on a genetic algorithm optimization is adopted, and the predicted computing power load demand of the user is taken as a limiting condition, all edge servers are collaboratively scheduled and distributed by a cloud center, and a final scheduling and distribution scheme is output.

[0032] Further, an alternating direction method of multipliers based on a genetic algorithm optimization specifically includes the following steps:

[0033] Step S41: Minimize the objective function design, calculate the running cost and network delay of the distributed edge server node and design the objective function:

[0034]

[0035] In the formula, representing the objective function, representing the running cost of the i-th edge server node, representing the network delay cost, representing the network delay of the i-th edge server node, representing the weight coefficient of the network delay;

[0036] Step S42: Constraint design, set the balance constraint of computing power supply and demand, node computing power limit, available computing power dynamic constraint and network delay constraint;

[0037] Step S43: Algorithm initialization, initialize the output computing power range of the edge server node, the genetic algorithm parameters and the alternating direction method of multipliers ADMM parameters, specifically including the following steps:

[0038] Step S431: Collect the real-time computing power parameters of the edge server node, and calculate the output computing power range of the corresponding edge server node;

[0039] Step S432: Initialize the population size, crossover probability, mutation probability and maximum number of generations of genetic algorithm, and randomly generate an initial population, each individual representing a set of scheduling scheme, that is, the output computing power allocated to each edge server node, wherein the output computing power of each set of scheduling scheme is greater than or equal to the predicted computing power load demand;

[0040] Step S433: Initialize ADMM parameters, including Lagrange multiplier, relaxation variable, step size parameter and ADMM convergence condition;

[0041] Step S44: Genetic algorithm optimization global search, genetic algorithm optimization is performed on the population, and the final optimal individual and the optimal scheduling scheme corresponding thereto are output, specifically including the following steps:

[0042] Step S441: Individual fitness calculation, for each individual in the population, decode the scheduling scheme corresponding to the output computing power of each edge server node and calculate the fitness function, that is, the objective function value of the population

[0043] Step S442: Individual selection, excellent individuals are selected according to the fitness value by using roulette wheel selection method; ​​

[0044] Step S443: excellent operation, cross operation is carried out to excellent individual, and new child population is generated;

[0045] Step S444: individual mutation, part of individual is randomly mutated and the fitness is recalculated;

[0046] Step S445: population update, the optimal individual is reserved according to the minimized fitness value, and the population is updated;

[0047] Step S446: when the maximum number of generations is reached, the genetic algorithm stops, the final optimal individual is obtained, and the optimal scheduling scheme corresponding to the final optimal individual is output as an initial solution into ADMM solving;

[0048] Step S45: ADMM local optimization and distributed scheduling optimization, and output the final scheduling allocation scheme, specifically comprising the following steps:

[0049] Step S451: local optimization problem decomposition, the minimized objective function is decomposed into N edge server nodes local optimization problems:

[0050] ;

[0051] In the formula, The sum of the running cost and network delay cost of the i-th edge server node in the optimal scheduling scheme is represented, The step length parameter is represented, The relaxation variable is represented, The Lagrange multiplier is represented;

[0052] Step S452: each edge server node independently calculates the local optimization problem, and updates the corresponding scheduling scheme, and obtains the updated scheduling scheme;

[0053] Step S453: the global relaxation variable is updated according to the updated scheduling scheme, and the updated global relaxation variable is obtained;

[0054] Step S454: the Lagrange multiplier is updated according to the updated scheduling scheme and the updated global relaxation variable;

[0055] Step S455: multiple rounds of optimization, repeat steps S451-S454 until the ADMM convergence condition is reached, and output the corresponding final scheduling allocation scheme.

[0056] The beneficial results obtained by the above scheme are as follows:

[0057] (1) The application adopts a space-time prediction algorithm based on multi-head attention and gate mechanism optimization, predicts and analyzes the terminal device computing power load, designs time encoding and space encoding, captures the time sequence characteristics of the computing power load and the spatial relationship between the edge server nodes, understands the interaction between the computing power load demands of each edge server node, and combines the multi-head attention mechanism with the expansion causal convolution, which can capture the instantaneous computing power load fluctuation and mine the long-term trend of the computing power load, so as to predict accurate computing power load and provide a reliable basis for subsequent computing power scheduling.

[0058] (2) In view of the problems of high network delay and poor computing power resource allocation in the prior art when processing edge node computing power, the application designs an alternating direction multiplier method based on genetic algorithm optimization, the genetic algorithm provides global search capability to avoid falling into local optimum, the ADMM alternating direction multiplier method realizes distributed decoupling to reduce the computational complexity, the genetic algorithm is suitable for nonlinear, multi-constraint optimization problems, and is compatible with multiple models, and the ADMM allows expansion to larger scale distributed edge node cloud computing system, the genetic algorithm quickly approaches the global optimal solution to improve the initial solution quality, speeds up the ADMM convergence, and improves the distributed collaborative allocation and scheduling efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of an edge cloud computing resource allocation optimization method based on deep learning provided by the application;

[0060] Figure 2 A flowchart of a space-time prediction algorithm based on multi-head attention and gate mechanism optimization;

[0061] Figure 3 A flowchart of an alternating direction multiplier method based on genetic algorithm optimization.

[0062] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0064] Embodiment one, refer to Figure 1 The application provides an edge cloud computing resource allocation optimization method based on deep learning, which specifically comprises the following steps:

[0065] Step S1: edge server deployment, distributed deployment of edge servers, wherein each edge server node can adjust its computing resources;

[0066] Step S2: communication and data sharing, connecting each edge server through Internet of Things technology to realize information interaction between nodes;

[0067] Step S3: computing power load prediction, using a spatio-temporal prediction algorithm based on multi-head attention and gate mechanism optimization to predict the computing power load demand of the terminal device;

[0068] Step S4: cloud distributed optimization and collaborative control, using an alternating direction multiplier method based on genetic algorithm optimization, to achieve the prediction of the user's computing power load demand as a limiting condition, through the cloud center to collaboratively schedule and distribute all edge servers, and output the final scheduling and distribution scheme.

[0069] Embodiment two, see Figure 2 , this embodiment is based on the above embodiment, step S3, specifically including the following steps:

[0070] Step S31: data collection, collecting the historical time series load demand M of the terminal device to each edge server, wherein the dimension of M is T x N x 1, T represents the step length of the historical time series, N represents the number of edge server nodes, and each node is collectively formed into a node graph according to the geographical position;

[0071] Step S32: spatio-temporal encoding, spatio-temporal feature encoding of the historical time series load demand M of each edge server node, to obtain the spatio-temporal encoding;

[0072] Step S33: dilated causal convolution processing, dilated causal convolution processing of the spatio-temporal encoding, to obtain the time series dependency relationship:

[0073] ;

[0074] In the formula, represents the value of the spatio-temporal encoding at time step t-k, represents the inflation rate, represents the convolution kernel weight, , that is, the time series dependency relationship;

[0075] Step S34: multi-head attention mechanism processing, multi-head attention processing of the spatio-temporal encoding and the time series dependency relationship, to obtain the spatio-temporal feature;

[0076] Step S35: feature fusion based on the gate mechanism, fusion optimization of the spatio-temporal feature and the time series dependency relationship based on the gate mechanism and residual connection, to obtain the optimized spatio-temporal feature:

[0077] ;

[0078] wherein, represents layer processing, represents layer normalization, represents spatio-temporal features, represents Sigmoid activation function, represents temporal dependency, represents spatio-temporal encoding, represents optimized spatio-temporal features;

[0079] Step S36: multi-layer perception processing and prediction output, performing multi-layer perception processing on the optimized spatio-temporal features and performing activation prediction to obtain predicted computing power load demand.

[0080] Embodiment three, refer to Figure 3 , based on the above embodiment, step S4, specifically comprising the following steps:

[0081] Step S41: design of the objective function for minimization, calculate the running cost and network delay of the distributed edge server nodes and design the objective function:

[0082] ;

[0083] wherein, represents the objective function, represents the running cost of the i-th edge server node, represents the network delay cost, represents the network delay of the i-th edge server node, represents the weight coefficient of the network delay;

[0084] Step S42: constraint condition design, set the balance constraint of computing power supply and demand, node computing power limit, available computing power dynamic constraint and network delay constraint;

[0085] Step S43: algorithm initialization, initialize the output computing power range of the edge server node, genetic algorithm parameters and alternating direction multiplier method ADMM parameters;

[0086] Step S44: genetic algorithm optimization global search, perform genetic algorithm optimization on the population and output the final optimal individual and the optimal scheduling scheme corresponding thereto;

[0087] Step S45: ADMM local optimization and distributed scheduling optimization, output the final scheduling allocation scheme, specifically comprising the following steps:

[0088] Step S451: local optimization problem decomposition, decompose the minimization objective function into N local optimization problems of edge server nodes:

[0089] ;

[0090] wherein, represents the sum of the running cost and network delay cost of the i-th edge server node in the optimal scheduling scheme, represents the step size parameter, represents the relaxation variable, represents the Lagrange multiplier;

[0091] Step S452: each edge server node independently calculates the local optimization problem and updates the corresponding scheduling scheme to obtain an updated scheduling scheme;

[0092] Step S453: update the global relaxation variable according to the updated scheduling scheme to obtain an updated global relaxation variable;

[0093] Step S454: update the Lagrange multiplier according to the updated scheduling scheme and the updated global relaxation variable;

[0094] Step S455: multiple rounds of optimization, repeat steps S451-S454 until the ADMM convergence condition is reached, and output the corresponding final scheduling allocation scheme.

[0095] In the fourth embodiment, the present scheme is used for real-time video analysis resource scheduling in intelligent transportation systems, and the method implementation steps are as follows:

[0096] Scenario description: a city intelligent transportation system deploys multiple edge server nodes to process high-definition video streams from traffic cameras for real-time vehicle recognition, traffic statistics, and violation detection tasks. Each edge node is responsible for video processing tasks in a certain area, and the task load changes dynamically over time (such as morning and evening rush hours) and space (such as road congestion);

[0097] Method implementation process:

[0098] Data collection and prediction (S3):

[0099] Collect historical video processing task loads (including frame rate, resolution, recognition model complexity, etc.) of each node;

[0100] Use a multi-head attention and gated mechanism spatio-temporal prediction model to predict the computing power demand of each node in the next 30 minutes;

[0101] Resource scheduling optimization (S4):

[0102] Objective function: minimize total processing delay and energy consumption;

[0103] Distributed scheduling using genetic algorithm + ADMM ensures that high-load nodes (such as main road cameras) have more computing power, and low-load nodes (such as night suburban intersections) are appropriately reduced in frequency to save energy;

[0104] Effect:

[0105] Prediction accuracy improved by 15%, resource utilization increased by 20%;

[0106] Task processing delay during peak hours decreased by 30%, overall system energy consumption decreased by 10%.

[0107] Example five, this scheme is used in industrial internet of things in the device collaborative computing, the method implementation steps are as follows:

[0108] Scenario description: A smart factory deploys multiple edge servers to process sensor data (such as temperature, vibration, image detection) for device state monitoring, anomaly detection, predictive maintenance, etc. The load of different production lines varies greatly, and there are sudden tasks (such as device fault detection);

[0109] Process of applying this method:

[0110] Data collection and prediction (S3):

[0111] Collect historical load data of each production line edge node;

[0112] Use space-time coding to capture spatial dependence between devices (such as adjacent device load correlation);

[0113] Resource scheduling optimization (S4):

[0114] Objective function: Minimize response time and resource waste;

[0115] Use GA + ADMM to dynamically allocate computing power, and prioritize critical devices (such as precision machining devices) for computing resources;

[0116] Effect:

[0117] Anomaly detection response time decreased from seconds to milliseconds;

[0118] Resource allocation is more balanced, and the ability to handle sudden tasks has improved by 25%.

[0119] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other present or future technologies can provide. Specifically, it is contemplated that some embodiments of the present application can be practiced without the specific details set forth in the preceding description.

[0120] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the appended claims and their equivalents.

[0121] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution should belong to the protection scope of the application.

Claims

1. A deep learning-based edge cloud computing resource allocation optimization method, characterized in that: Specifically, the following steps are included: Step S1: Edge server deployment. Edge servers are deployed in a distributed manner, where each edge server node can adjust its own computing resources. Step S2: Communication and data sharing, connecting various edge servers through IoT technology to achieve information interaction between nodes; Step S3: Computing load prediction. A spatiotemporal prediction algorithm based on multi-head attention and gating mechanism optimization is used to predict the computing load demand of terminal devices. Specifically, it includes the following steps: Step S31: Data collection. Collect the historical time-series load requirements M of the terminal device for each edge server. M has a dimension of T×N×1, where T represents the step size of the historical time series, N represents the number of edge server nodes, and each node forms a node graph according to its geographical location. Step S32: Spatiotemporal coding. Spatiotemporal features are encoded for the historical time-series load demand M of each edge server node to obtain the spatiotemporal code. This specifically includes the following steps: Step S321: Time encoding, performing time encoding on the historical time series: ； ； In the formula, The step size representing the historical sequence, Represents the dimension of the hidden layer in the encoding. Representing a moment in the historical timeline, and This is time encoding; Step S322: Global structure information acquisition. The node graph is spatially encoded using a normalized Laplacian matrix to obtain the global structure matrix. Step S323: Spatial coding, perform matrix multiplication on the global structure matrix and the historical time-series load demand M to obtain the spatial code; Step S324: Spatiotemporal coding, fusing spatial coding and temporal coding to obtain spatiotemporal coding; Step S33: Dilated causal convolution processing: Perform dilated causal convolution processing on the spatiotemporal encoding to obtain the temporal dependencies: ； In the formula, This represents the value of the spatiotemporal code at time step tk. Represents the expansion rate. Represents the convolution kernel weights. Represents the kernel size. This represents a temporal dependency; Step S34: Multi-head attention mechanism processing, performing multi-head attention processing on spatiotemporal encoding and temporal dependencies to obtain spatiotemporal features, specifically including the following steps: Step S341: Generate query vector and key vector by performing a linear transformation on the spatiotemporal encoding to generate query vector and key vector; Step S342: Value vector generation, linearly encoding the temporal dependency to generate a value vector; Step S343: Apply attention mechanism to the query vector, key vector, and value vector: ； In the formula, These represent the query vector, key vector, and value vector, respectively. Represents the dimensions of the query vector and the key vector. Represents the activation function; Step S344: Multi-head attention optimization, perform multi-head attention optimization on the results of the attention mechanism processing to obtain spatiotemporal features; Step S35: Feature fusion based on gating mechanism: The spatiotemporal features and temporal dependencies are fused and optimized using gating mechanism and residual connections to obtain optimized spatiotemporal features. ； In the formula, represent Layer processing, Representative level normalization, Represents spatiotemporal characteristics, Represents the Sigmoid activation function. Represents temporal dependency. Represents spatiotemporal coding. Represents optimized spatiotemporal features; Step S36: Multilayer perceptron processing and prediction output: The optimized spatiotemporal features are processed by a multilayer perceptron and activated for prediction to obtain the predicted computing load requirements. Step S4: Cloud-based distributed optimization and collaborative control. An alternating direction multiplier method based on genetic algorithm optimization is adopted to predict the user's computing power load demand as a constraint. The cloud center performs collaborative scheduling and allocation of all edge servers and outputs the final scheduling and allocation scheme. Furthermore, an alternating direction multiplier method based on genetic algorithm optimization specifically includes the following steps: Step S41: Minimize the objective function design, calculate the operating cost and network latency of the distributed edge server nodes, and design the objective function: ； In the formula, Represents the objective function. This represents the operating cost of the i-th edge server node. Represents network latency cost, This represents the network latency of the i-th edge server node. Weighting coefficients representing network latency; Step S42: Design constraints, set computing power supply and demand balance constraints, node computing power limits, available computing power dynamic constraints, and network latency constraints; Step S43: Algorithm initialization. Initialize the output computing power range of the edge server nodes, the genetic algorithm parameters, and the alternating direction multiplier method (ADMM) parameters. This includes the following steps: Step S431: Collect the computing power parameters of the real-time edge server nodes and calculate the output computing power range of the corresponding edge server nodes; Step S432: Initialize the population size, crossover probability, mutation probability, and maximum number of generations of inheritance, and randomly generate an initial population. Each individual represents a set of scheduling schemes, i.e., the output computing power allocated to each edge server node. The output computing power of each set of scheduling schemes is greater than or equal to the predicted computing power load requirement. Step S433: Initialize ADMM parameters, including Lagrange multipliers, slack variables, step size parameters, and ADMM convergence conditions; Step S44: Genetic algorithm optimizes the global search. The population is optimized using a genetic algorithm, and the final optimal individual and its corresponding optimal scheduling scheme are output. This includes the following steps: Step S441: Individual fitness calculation. For each individual in the population, decode the scheduling scheme corresponding to the output computing power of each edge server node and calculate the fitness function, i.e., the objective function value of the population. ; Step S442: Individual selection, using roulette wheel selection to select superior individuals based on fitness values; Step S443: Excellent operation, perform crossover operation on excellent individuals to generate a new offspring population; Step S444: Individual variation, randomly mutate a subset of individuals and recalculate fitness; Step S445: Population update: Retain the best individual based on the minimized fitness value and update the population; Step S446: When the maximum number of generations of inheritance is reached, the genetic algorithm stops, obtains the final optimal individual, and outputs its corresponding optimal scheduling scheme as the initial solution to be solved in ADMM. Step S45: ADMM local optimization and distributed scheduling optimization, outputting the final scheduling allocation scheme, specifically including the following steps: Step S451: Decompose the local optimization problem, decompose the minimization objective function into a local optimization problem for N edge server nodes: ； In the formula, This represents the sum of the operating cost and network latency cost of the i-th edge server node in the optimal scheduling scheme. Represents the step size parameter. Represents slack variables. Represents the Lagrange multipliers; Step S452: Each edge server node independently calculates the local optimization problem and updates the corresponding scheduling scheme to obtain the updated scheduling scheme; Step S453: Update the global relaxation variables according to the updated scheduling scheme to obtain the updated global relaxation variables; Step S454: Update the Lagrange multipliers according to the updated scheduling scheme and the updated global relaxation variables; Step S455: Perform multiple rounds of optimization, repeating steps S451-S454 until the ADMM convergence condition is met, and output the corresponding final scheduling allocation scheme.

Images