MEC network statistical isomerism analysis and dynamic resource allocation method under isomorphic terminals

Abstract

The invention discloses an MEC network statistical isomerism analysis and dynamic resource allocation method under an isomorphic terminal, and relates to the technical field of edge computing, and the method comprises the following steps: collecting the size of an isomorphic terminal sample data set and local training time through a log and a monitoring tool, carrying out the power law distribution classification after preprocessing, setting a threshold value classification device, and carrying out the classification of the local training time; lSTM is introduced to predict variance trend, samples and iteration times are intelligently adjusted, local training is optimized, delay and data volume weight are fused, and asynchronous updating transmission of a global model is realized. According to the method, effective analysis of MEC network statistical heterogeneity under the isomorphic terminal is realized through refined data acquisition and preprocessing in combination with an LSTM prediction technology, the flexibility and efficiency of resource allocation are remarkably improved, meanwhile, a global model updating strategy fusing delay and data volume weight ensures timeliness and accuracy of model updating, and the method has the advantages of being high in practicability and easy to popularize. And powerful support is provided for edge calculation application.

Description

Technical Field

[0001] The present invention relates to the technical field of edge computing, and particularly to an analysis method for statistical heterogeneity and a dynamic resource allocation method of an MEC network under homogeneous terminals. Background Art

[0002] The MEC network, namely the multi-access edge computing network, is an emerging technical paradigm. It aims to move computing resources and capabilities from centralized data centers to the edge of the network, closer to data sources and end-users, in order to reduce network latency and improve service quality. The core idea of the MEC network is to deploy computing and storage capabilities at the network edge to provide applications and services with low latency, high bandwidth, and real-time processing for end-users. Specifically, the MEC network is built on mobile network infrastructure and can utilize its physical proximity to users to provide low-latency and highly reliable services. In addition, the MEC network can also provide context-related information for devices and applications, enabling more intelligent and efficient decision-making. This makes the MEC network have broad application prospects in scenarios such as the Internet of Things, 5G communication, autonomous driving, and smart cities.

[0003] To solve the problems existing in the MEC network in terms of data heterogeneity and dynamic resource allocation, the existing technology mainly adopts resource allocation strategies based on traditional network architectures, that is, network resources are allocated according to preset rules and algorithms. However, when facing the complex and changeable data heterogeneity of the MEC network, this method often results in unreasonable resource allocation and low efficiency. Due to data heterogeneity, the resource requirements of different terminals and services vary significantly, and traditional resource allocation methods are difficult to accurately capture these differences, thus leading to problems such as resource waste and service quality degradation. To overcome this problem, an analysis method for statistical heterogeneity and a dynamic resource allocation method of an MEC network under homogeneous terminals are proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an analysis method for statistical heterogeneity and a dynamic resource allocation method of an MEC network under homogeneous terminals to solve the problems raised in the above background art.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: an analysis method for statistical heterogeneity and a dynamic resource allocation method of an MEC network under homogeneous terminals, including the following steps:

[0006] S1. Through device logs and storage monitoring tools, collect the size of the sample data set of homogeneous terminal devices, and collect the local update and training time of the terminal devices through performance monitoring tools and timers;

[0007] S2. Preprocess the collected local data, set the threshold of the local sample data volume through power-law distribution classification technology, and classify the devices;

[0008] S3. Introduce the LSTM model to predict the variance change trend among different batches of samples, intelligently adjust the sample size, use the dynamic sampling algorithm to calculate the dynamic sample size of large-sample devices, achieve local training optimization driven by gradient variance, and dynamically adjust the number of iterations according to the variance trend;

[0009] S4. Integrate delay weighting and data volume weights to optimize the global model update, and through the wireless channel and the scheduling thread, achieve the global model distribution between the server and the terminal device and the asynchronous transmission of updated parameters.

[0010] A further improvement of the technical solution of the present invention is that in S1, the process of collecting the sample data set size of homogeneous terminal devices through device logs and storage monitoring tools, and collecting the local update training time of terminal devices through performance monitoring tools and timers includes:

[0011] Deploy the terminal devices in the same network environment and connect them to the MEC server. The terminal devices include smart phones, tablets and sensor nodes. Configure log recording tools, performance monitoring tools and timers on the terminal devices;

[0012] The log recording tool regularly scans the storage directory on the terminal device, obtains the file sizes in the directory, accumulates the sizes of relevant files, and obtains the total sample data set size , the performance monitoring tool starts the timer before the start of each round of local model training, stops the timer after the training ends, records the difference between the two timestamps as the local update training time for this time, and reports the local sample data set size and local update training time data to the MEC server before each iteration.

[0013] A further improvement of the technical solution of the present invention is that in S2, the process of preprocessing the collected sample data set size and local update training time includes:

[0014] Adopt the 3σ principle to remove the noise data in the local sample data set size and local update training time, use the linear interpolation method to fill the missing values in the local sample data set, record the local update training time of each terminal device, check whether it is within a reasonable range, if there are data points deviating from the expected value, then use statistical methods to identify and remove the outliers, and fill the corresponding missing data with the mean value of the local update training time of the terminal device.

[0015] A further improvement of the technical solution of the present invention is that in S2, the process of setting the local sample data volume threshold and classifying devices through the power-law distribution classification technology includes:

[0016] Based on the Pareto principle, set the sample data volume size threshold to Calculate the cumulative distribution function of the dataset size and set as the 80th percentile of the sample dataset size. If , then the device belongs to a large sample. If , then the device belongs to a small sample.

[0017] A further improvement of the technical solution of the present invention lies in that: in the step S3, the process of introducing an LSTM model to predict the variance change trend between different batches of samples includes:

[0018] Train the LSTM model using historical data. Input a dataset containing the preprocessed historical sample variance estimates, the sample dataset size, and the local update training time, so that the sample dataset size and the local update training time included in each batch form a two-dimensional array. Calculate the sample variance of each batch, define the number of time steps and the number of features, and convert the historical data into a form suitable for input to the LSTM model. Use every n consecutive time step data as samples to generate an input matrix and a target vector. Construct a neural network model including a single-layer LSTM unit and a fully connected layer. The LSTM unit captures the long-term and short-term dependencies in the time series, and the fully connected layer outputs the predicted value of the sample variance of the next batch. Minimize the mean square error between the predicted value of the sample variance and the actual value, and optimize the model parameters through the backpropagation algorithm.

[0019] A further improvement of the technical solution of the present invention lies in that: in the step S3, the process of intelligently adjusting the sample size using the trained LSTM model includes:

[0020] During the process of dynamically adjusting the sample size, set the maximum sample size. Use the trained LSTM model to predict the sample variance of the next batch. Based on the predicted sample variance and the set variance threshold, dynamically adjust the sample size of the current batch;

[0021] Set the predicted variance threshold. If the predicted sample variance is higher than the set threshold, it is considered that the data distribution is heterogeneous. Increase the computing resources, extend the training time, reduce the number of parallel tasks, and for large sample devices, use a dynamic sampling algorithm to gradually increase the sample size until the best convergence point is reached. If the predicted sample variance is lower than the set threshold, it is considered that the data distribution is relatively uniform. Reduce the computing resources, shorten the training time, increase the number of parallel tasks, and for small sample devices, use all samples for training. If the predicted sample variance is close to the threshold, maintain the current computing resource allocation and training time settings, and strengthen the monitoring of the MEC network status.

[0022] A further improvement of the technical solution of the present invention lies in that: in the step S3, the process of calculating the dynamic sample size of a large sample device using a dynamic sampling algorithm to achieve local training optimization driven by gradient variance includes:

[0023] An initial sample size is selected for each large-sample device. In each iteration, the terminal device calculates the gradient using the current sample size and estimates its variance. According to the change of the gradient variance, the sample size in the next iteration is adaptively adjusted. If the gradient variance under the current sample size is large, the sample size is increased to reduce the variance. If the gradient variance under the current sample size is small, the sample size is reduced to improve the efficiency;

[0024] In each iteration, gradient descent calculation is performed based on the adjusted sample size to optimize local training. The local model parameters of the terminal device are updated according to the calculated gradient, and the updated model parameters are uploaded to the MEC server.

[0025] A further improvement of the technical solution of the present invention is that in S3, the process of dynamically adjusting the number of iterations according to the variance trend includes:

[0026] Monitor the change rate of the loss function in each round of training to adjust the remaining number of iterations of each terminal device. Set the global maximum number of iterations, and calculate the initial number of iterations according to the average value of the local sample data set sizes of each device , based on the adjusted sample size, calculate the number of iterations of each terminal device, set the number of iterations adjusted each time, perform gradient descent calculation on the selected samples to obtain the initial gradient variance. If the gradient variance under the current sample size is large, increase the number of iterations to reduce the variance. If the gradient variance under the current sample size is small, reduce the number of iterations to improve the efficiency, where, represents the average value of the sample data set sizes collected by the terminal device.

[0027] A further improvement of the technical solution of the present invention is that in S4, the process of fusing delay weighting and data volume weight to optimize the global model update includes:

[0028] The terminal device is connected to the MEC server through a wireless network. A scheduling thread is set on the MEC server to manage the upload and download tasks of each terminal device, monitor the transmission delay and the size of the sample data volume of each terminal device, set the delay weighting factor and the data volume weight, and calculate the comprehensive weight.

[0029] A further improvement of the technical solution of the present invention is that in S4, the process of realizing the asynchronous transmission of the global model distribution and the updated parameters between the server and the terminal device through the wireless channel and the scheduling thread includes:

[0030] The asynchronous communication mechanism enables the device to upload the updated model parameters after local training at any time by scheduling threads and update threads to execute in parallel, without waiting for other devices to synchronize and complete. The scheduling thread is responsible for periodically sending the globally shared model and timestamp to the terminal device to trigger the local training task of the device. The update thread receives the local model parameters uploaded by each device and updates the global model. If the device cannot complete the training on time due to resource constraints, the server does not wait for this device and continues to process the feedback from other devices.

[0031] Due to the adoption of the above technical solutions, the technical progress achieved by the present invention compared with the prior art is as follows:

[0032] 1. The present invention provides a method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals. By accurately collecting and analyzing the sample dataset size and local update training time of homogeneous terminal devices, in-depth insight into the performance of terminal devices is achieved. This method effectively identifies data heterogeneity, provides accurate data support for subsequent resource allocation, significantly improves resource utilization efficiency, and reduces resource waste.

[0033] 2. The present invention provides a method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals. It innovatively introduces the LSTM model to predict the variance change trend between samples, and intelligently adjusts the sample size and iteration times, realizing local training optimization driven by gradient variance. This innovation not only improves the training efficiency but also ensures the stability and accuracy of model training, providing new ideas for dynamic resource allocation in the MEC network.

[0034] 3. The present invention provides a method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals. It combines delay weighting and data volume weighting to optimize the global model update strategy. Through the asynchronous transmission of the wireless channel and the scheduling thread, efficient communication between the server and the terminal device is achieved, ensuring the timeliness and accuracy of global model updates. This method not only improves the overall performance of the MEC network but also provides strong support for application scenarios with high real-time requirements. Brief Description of the Drawings

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0036] Figure 1 It is a flowchart of the present invention. Detailed Embodiments

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] An embodiment is as Figure 1 shown. The present invention provides a method for MEC network statistical heterogeneity analysis and dynamic resource allocation under homogeneous terminals, including the following steps:

[0039] S1. Through device logs and storage monitoring tools, collect the size of the sample data set of homogeneous terminal devices. Through performance monitoring tools and timers, collect the local update training time of the terminal devices. Deploy the terminal devices in the same network environment and connect them to the MEC server. The terminal devices include smart phones, tablets, and sensor nodes. Configure log recording tools, performance monitoring tools, and timers on the terminal devices. The log recording tool regularly scans the storage directory on the terminal device, obtains the file sizes in the directory, accumulates the sizes of relevant files, and obtains the total sample data set size. The performance monitoring tool starts the timer before each round of local model training and stops the timer after the training ends, records the difference between the two timestamps as the local update training time for this time, and reports the local sample data set size and local update training time data to the MEC server before each iteration.

[0040] S2. Preprocess the collected local data. Through power-law distribution classification technology, set the local sample data volume threshold and classify the devices. Use the 3σ principle to remove the noise data in the local sample data set size and local update training time, use linear interpolation to fill the missing values in the local sample data set, record the local update training time of each terminal device, and check whether it is within a reasonable range. If there are data points deviating from the expected value, use statistical methods to identify and remove the outliers, and fill the corresponding missing data with the mean of the local update training time of the terminal device. Based on the Pareto principle, set the sample data volume size threshold as Calculate the cumulative distribution function of the data set size, and set as the 80th percentile of the sample data set size. If , then the device belongs to a large sample. If , then the device belongs to a small sample.

[0041] S3. Introduce the LSTM model to predict the variance change trend among different batches of samples, intelligently adjust the sample size, use the dynamic sampling algorithm to calculate the dynamic sample size of large-sample devices, achieve local training optimization driven by gradient variance, dynamically adjust the number of iterations according to the variance trend, train the LSTM model using historical data, input a dataset containing the preprocessed historical sample variance estimates, the size of the sample dataset, and the local update training time, form a two-dimensional array with the size of the sample dataset and the local update training time included in each batch, calculate the sample variance of each batch, define the number of time steps and the number of features, and convert the historical data into a form suitable for input to the LSTM model. Take every n consecutive time-step data as samples to generate the input matrix and the target vector. Build a neural network model containing a single-layer LSTM unit and a fully connected layer. The LSTM unit captures the long-term and short-term dependencies in the time series, and the fully connected layer outputs the predicted value of the sample variance for the next batch. Minimize the mean squared error between the predicted value of the sample variance and the actual value, and optimize the model parameters through the backpropagation algorithm. During the process of dynamically adjusting the sample size, set the maximum sample size, use the trained LSTM model to predict the sample variance of the next batch, and based on the predicted sample variance and the set variance threshold, dynamically adjust the sample size of the current batch. Set the predicted variance threshold. If the predicted sample variance is higher than the set threshold, it is considered that the data distribution is heterogeneous, increase the computing resources, extend the training time, reduce the number of parallel tasks, and large-sample devices use the dynamic sampling algorithm to gradually increase the sample size until the best convergence point is reached. If the predicted sample variance is lower than the set threshold, it is considered that the data distribution is relatively uniform, reduce the computing resources, shorten the training time, increase the number of parallel tasks, and small-sample devices use all samples for training. If the predicted sample variance is close to the threshold, maintain the current computing resource allocation and training time settings, strengthen the monitoring of the MEC network status, select the initial sample size for each large-sample device. In each iteration, the terminal device calculates the gradient and estimates its variance using the current sample size. According to the change of the gradient variance, adaptively adjust the sample size in the next iteration. If the gradient variance under the current sample size is large, increase the sample size to reduce the variance. If the gradient variance under the current sample size is small, reduce the sample size to improve efficiency. In each iteration, perform gradient descent calculation based on the adjusted sample size to optimize local training, update the local model parameters of the terminal device according to the calculated gradient, upload the updated model parameters to the MEC server, monitor the change rate of the loss function in each round of training to adjust the remaining number of iterations for each terminal device, set the global maximum number of iterations, and calculate the initial number of iterations according to the average value of the local sample dataset sizes of each device , based on the adjusted sample size, calculate the number of iterations for each terminal device, set the number of iterations for each adjustment, perform gradient descent calculation on the selected samples to obtain the initial gradient variance. If the gradient variance under the current sample size is large, increase the number of iterations to reduce the variance; if the gradient variance under the current sample size is small, reduce the number of iterations to improve efficiency, where represents the average value of the sample data set size of the acquisition terminal device;

[0042] S4. Integrate the delay weight and data volume weight to optimize the global model update. Through the wireless channel and the scheduling thread, realize the asynchronous transmission of the global model distribution and updated parameters between the server and the terminal device. The terminal device is connected to the MEC server through the wireless network. Set a scheduling thread on the MEC server to manage the upload and download tasks of each terminal device, monitor the transmission delay and sample data volume size of each terminal device, set the delay weight factor and data volume weight, calculate the comprehensive weight, and the asynchronous communication mechanism is executed in parallel through the scheduling thread and the update thread, enabling the device to upload the updated model parameters after local training at any time without waiting for other devices to synchronize and complete. The scheduling thread is responsible for regularly sending the global shared model and timestamp to the terminal device to trigger the local training task of the device. The update thread receives the local model parameters uploaded by each device and updates the global model. If the device fails to complete the training on time due to resource constraints, the server does not wait for this device and continues to process the feedback from other devices.

[0043] As mentioned above, it is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. Analysis method for statistical heterogeneity and dynamic resource allocation method in MEC network under isomorphic terminals, characterized in that, It includes the following steps: S1. Through device logs and storage monitoring tools, collect the size of the homogeneous terminal device sample data set, and through performance monitoring tools and timers, collect the local update training time of the terminal device; S2. Preprocess the collected local data, set the local sample data volume threshold through power-law distribution classification technology, and classify the devices; S3. Introduce the LSTM model to predict the variance change trend between different batches of samples, intelligently adjust the sample size, use the dynamic sampling algorithm to calculate the dynamic sample volume of large-sample devices, achieve local training optimization driven by gradient variance, and dynamically adjust the number of iterations according to the variance trend; S4. Integrate delay weighting and data volume weights to optimize the global model update, and through the wireless channel and scheduling thread, achieve the global model distribution between the server and the terminal device and the asynchronous transmission of updated parameters.

2. The method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals according to claim 1, wherein: In the above S1, the process of collecting the size of the homogeneous terminal device sample data set through device logs and storage monitoring tools and collecting the local update training time of the terminal device through performance monitoring tools and timers includes: Deploy the terminal devices in the same network environment and connect them to the MEC server. The terminal devices include smart phones, tablets, and sensor nodes. Configure log recording tools, performance monitoring tools, and timers on the terminal devices; The logging tool periodically scans the storage directory on the terminal device, obtains the file sizes in the directory, accumulates the sizes of relevant files, and gets the total size of the sample data set. , Before each round of local model training starts, the performance monitoring tool starts a timer, stops the timer after the training ends, records the difference between the two timestamps as the local update training time for this time, and reports the local sample data set size and local update training time data to the MEC server before each iteration.

3. The method for analyzing the statistical heterogeneity and dynamic resource allocation of the MEC network under homogeneous terminals according to claim 2, wherein: In the above S2, the process of preprocessing the collected sample data set size and local update training time includes: Adopt the 3σ principle to remove the noise data in the local sample data set size and local update training time, use linear interpolation to fill in the missing values in the local sample data set, record the local update training time of each terminal device, check whether it is within a reasonable range. If there are data points deviating from the expected value, use statistical methods to identify and remove the outliers, and fill in the corresponding missing data with the mean value of the local update training time of the terminal device.

4. The method for analyzing the statistical heterogeneity and dynamic resource allocation of the MEC network under homogeneous terminals according to claim 3, wherein: In the above S2, the process of setting the local sample data volume threshold and classifying the devices through power-law distribution classification technology includes: Set the threshold of the sample data volume size based on the Pareto principle to be , calculate the cumulative distribution function of the data set size, and set to be the 80th percentile of the sample data set size. If , then the device belongs to a large sample. If , then the device belongs to a small sample.

5. The method for analyzing statistical heterogeneity and dynamically allocating resources in an MEC network under homogeneous terminals according to claim 4, wherein: In the above S3, the process of introducing the LSTM model to predict the variance change trend between different batches of samples includes: Train the LSTM model with historical data. Input a data set containing the preprocessed historical sample variance estimation values, sample data set size, and local update training time, so that the sample data set size and local update training time included in each batch form a two-dimensional array, calculate the sample variance of each batch, define the number of time steps and the number of features, and convert the historical data into a form suitable for the input of the LSTM model. Use every n consecutive time step data as a sample to generate an input matrix and a target vector. Build a neural network model containing a single-layer LSTM unit and a fully connected layer. The LSTM unit captures the long-term and short-term dependencies in the time series, and the fully connected layer outputs the predicted value of the sample variance of the next batch. Minimize the mean square error between the predicted value of the sample variance and the actual value, and optimize the model parameters through the backpropagation algorithm.

6. The method for analyzing statistical heterogeneity and dynamically allocating resources in an MEC network under homogeneous terminals according to claim 5, wherein: In the above S3, the process of intelligently adjusting the sample size using the trained LSTM model includes: During the process of dynamically adjusting the sample size, set the maximum sample size, use the trained LSTM model to predict the sample variance of the next batch, and based on the predicted sample variance and the set variance threshold, dynamically adjust the sample size of the current batch; Set the predicted variance threshold. If the predicted sample variance is higher than the set threshold, it is considered that the data distribution is heterogeneous, increase computing resources, extend the training time, reduce the number of parallel tasks, and for large-sample devices, use the dynamic sampling algorithm to gradually increase the sample size until the optimal convergence point is reached. If the predicted sample variance is lower than the set threshold, it is considered that the data distribution is relatively uniform, reduce computing resources, shorten the training time, increase the number of parallel tasks, and for small-sample devices, use all samples for training. If the predicted sample variance is close to the threshold, maintain the current computing resource allocation and training time settings, and strengthen the monitoring of the MEC network status.

7. The method for MEC network statistical heterogeneity analysis and dynamic resource allocation under homogeneous terminals according to claim 6, characterized in that: In S3, the process of using the dynamic sampling algorithm to calculate the dynamic sample size of large-sample devices and realizing local training optimization driven by gradient variance includes: Select an initial sample size for each large-sample device. In each iteration, the terminal device uses the current sample size to calculate the gradient and estimate its variance. According to the change of the gradient variance, adaptively adjust the sample size in the next iteration. If the gradient variance under the current sample size is large, increase the sample size to reduce the variance. If the gradient variance under the current sample size is small, reduce the sample size to improve efficiency; In each iteration, perform gradient descent calculation based on the adjusted sample size to optimize local training, update the local model parameters of the terminal device according to the calculated gradient, and upload the updated model parameters to the MEC server.

8. The method for analyzing statistical heterogeneity and dynamic resource allocation in the MEC network under homogeneous terminals according to claim 7, wherein: In S3, the process of dynamically adjusting the number of iterations according to the variance trend includes: Monitor the change rate of the loss function in each round of training to adjust the remaining number of iterations of each terminal device. Set the global maximum number of iterations, and calculate the initial number of iterations based on the average of the local sample dataset sizes of each device. , calculate the number of iterations of each terminal device based on the adjusted sample size, set the number of iterations for each adjustment, perform gradient descent calculation on the selected samples to obtain the initial gradient variance. If the gradient variance under the current sample size is large, increase the number of iterations to reduce the variance; if the gradient variance under the current sample size is small, reduce the number of iterations to improve efficiency, where represents the average of the sample dataset sizes of the acquisition terminal devices.

9. The method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals according to claim 8, wherein: In S4, the process of fusing delay weighting and data volume weight to optimize the global model update includes: The terminal device is connected to the MEC server through a wireless network. Set a scheduling thread on the MEC server to manage the upload and download tasks of each terminal device, monitor the transmission delay and sample data volume size of each terminal device, set the delay weighting factor and data volume weight, and calculate the comprehensive weight.

10. The method for analyzing statistical heterogeneity and dynamic resource allocation in an MEC network under homogeneous terminals according to claim 9, wherein: In S4, the process of realizing the asynchronous transmission of the global model distribution and updated parameters between the server and the terminal device through the wireless channel and the scheduling thread includes: The asynchronous communication mechanism is executed in parallel through the scheduling thread and the update thread, enabling the device to upload the updated model parameters at any time after completing local training without waiting for other devices to synchronize. The scheduling thread is responsible for regularly sending the global shared model and timestamp to the terminal device to trigger the local training task of the device. The update thread receives the local model parameters uploaded by each device and updates the global model. If the device cannot complete the training on time due to resource constraints, the server does not wait for this device and continues to process the feedback of other devices.

Images