An application-level traffic prediction and model migration method at the network edge

Abstract

The invention discloses an application-level traffic prediction and model migration method for network edge, comprising: determining a model migration domain, using an unsupervised clustering algorithm based on application traffic statistical characteristics to complete clustering of edge nodes; selecting a migration model, wherein the migration model is a universal model selected in the migration domain and migrated to each edge computing node; classifying the traffic sequences of multiple network applications under each edge computing node based on the similarity in time and shape distribution according to the application clustering algorithm; after classifying multiple network applications under each edge node, N categories are obtained, each category trains its own multi-application traffic prediction model, generates N multi-application traffic prediction models, and saves these models to wait for edge node calls; when the N multi-application traffic prediction models are trained, they are called back by the edge computing node; in the migration domain, the N multi-application traffic prediction models are respectively migrated to all nodes in multiple domains for deployment and application.

Description

Technical Field

[0001] The present invention relates to the fields of edge computing, network traffic prediction, and edge model migration. Specifically, the present invention relates to an application-level traffic prediction and model migration solution for the network edge. Background Art

[0002] Edge computing sinks certain computing and storage resources to the edge of the network, providing high bandwidth and low latency service guarantees, thereby enhancing the user experience. Application-level traffic prediction can provide the ability to predict multiple Internet applications at the same time, thereby helping operators provide fine-grained application-level network management and application-level optimization capabilities. Therefore, carrying out application-level traffic prediction at the edge of the network can perceive fine-grained network traffic changes in a more timely, dynamic, and intelligent manner, thereby serving network resource allocation, network performance optimization, green network development and other network development needs. At the same time, with the deployment and application of edge computing nodes in the future, it is unrealistic to maintain a traffic prediction model for each network application under the coverage of each edge node, and it will also bring huge data collection and processing and model training overhead, which is not in line with the development outline of the green network and will also increase the operating costs of operators.

[0003] Application-level prediction of traffic from the location node where network traffic is generated is an important guarantee for fine-grained and timely prediction, and the network edge is the best deployment location. At the same time, in order to reduce the overhead of maintaining too many models, guiding the migration of prediction models at the edge of the network based on the similarity of network traffic distribution under edge nodes is an important measure to reduce overhead. The existing model only focuses on the prediction of the total regional traffic volume, and the model migration considers the interaction between the edge and the cloud. However, under the paradigm of edge computing in the future, more than 70% of the traffic will be unloaded and processed at the edge, so model migration between edge nodes is the main application requirement in the future. In view of the above application scenarios, the present invention proposes an application-level traffic prediction and model migration solution at the edge of the network. Summary of the invention

[0004] The purpose of the present invention is to solve the above-mentioned problems existing in the existing prediction models.

[0005] To achieve the above object, the present invention provides an application-level traffic prediction and model migration method at the edge of a network, the method comprising the following steps:

[0006] Determine the model migration domain and use an unsupervised clustering algorithm based on application traffic statistical characteristics to complete the clustering of edge nodes;

[0007] Select a migration model, where the migration model is a universal model selected in the migration domain to be migrated to each edge computing node;

[0008] The traffic sequences of various network applications under each edge computing node are classified based on their similarity in time and shape distribution according to the application clustering algorithm;

[0009] After classifying multiple network applications under each edge node, N categories are obtained. Each category trains its own multi-application traffic prediction model, and finally generates N multi-application traffic prediction models. These models are saved and wait for edge nodes to call them.

[0010] When the training of N multi-application traffic prediction models is completed, they are called back by the edge computing node;

[0011] In the migration domain, N multi-application traffic prediction models are migrated to all nodes in multiple domains for deployment and application.

[0012] The present invention is deployed on edge computing nodes to predict the network traffic of multiple applications in the region, and at the same time completes the migration and deployment of the model in the migration domain through the model migration strategy, thereby completing fine-grained, application-level network traffic prediction on the basis of extremely low overhead to serve and dynamically allocate network resources, optimize network applications and develop green networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a flow chart of an application-level traffic prediction and model migration method for a network edge provided by an embodiment of the present invention;

[0014] Figure 2 for Figure 1 The architecture diagram of the application-level traffic prediction and model migration method at the network edge is shown;

[0015] Figure 3 for Figure 1 Architecture diagram of the multi-application traffic prediction model in the method shown. DETAILED DESCRIPTION

[0016] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the above description. Although the present embodiment shows exemplary embodiments of the present disclosure, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to enable the scope of the present disclosure to be fully communicated to those skilled in the art.

[0017] The embodiments of the present invention are deployed on edge computing nodes in various regions. Each edge computing node contains all library files required by the prediction model, and no additional overhead is required for model deployment and migration.

[0018] Figure 1 A schematic diagram of a flow chart of a network edge application-level traffic prediction and model migration method provided by an embodiment of the present invention. Figure 1 As shown, the method includes steps S101-S106:

[0019] Step S101, determine the model migration domain: use an unsupervised clustering algorithm based on application traffic statistical characteristics to complete the clustering of edge nodes. For each application, the four statistical features of the traffic sequence [Max, Min, Mean, Std] are used. Then for N applications, the dimension of the input matrix of the K-means algorithm is [4*N]. The clustering result determines the optimal number of clusters based on the silhouette score (silhouette coefficient), and its value range is [-1,1]. The larger the value, the better the clustering result. The edge nodes in each cluster show similarity in the statistical characteristics of multi-application traffic.

[0020] Step S102, migration model selection: The migration model refers to selecting a universal model that can be migrated to each edge computing node in the migration domain, which generates a model trained by the traffic of multiple applications and the largest node in the migration domain. The model of the node with a large amount of data can learn more time series patterns and can be directly migrated to similar nodes in the migration domain.

[0021] Step S103, application clustering: An edge computing node may include a variety of network applications, such as streaming media, conference office, chat, file transfer, etc. These applications have great differences in network bandwidth requirements and latency requirements. Therefore, there are fundamental differences in the time series distribution of traffic. Application clustering is an application clustering algorithm proposed according to an embodiment of the present invention to classify traffic sequences according to the similarity in time and shape distribution. The similarity between applications is determined based on the Wasserstein distance. Compared with the traditional Euclidean distance, this distance can take into account the similarity of the shape distribution between two sequences, thereby more accurately classifying applications to reduce the information entropy within the class, thereby reducing the problem of under-fitting of the model.

[0022] S104, model aggregation: for each edge node, multiple network applications are divided into N categories after step S103, each category trains its own multi-application traffic prediction model, and finally generates N traffic prediction models, which are saved and wait for the edge node to call.

[0023] S105, model callback: After the training of N multi-application prediction models is completed, they are called back by the edge computing node. These N models are all generated by the node with the largest traffic in the migration domain.

[0024] S106, model migration: within the migration domain, N models are migrated to all nodes in multiple domains for deployment and application.

[0025] Figure 2 for Figure 1 The architecture diagram of the application-level traffic prediction and model migration method at the network edge is shown in FIG. Figure 2 As shown in the figure, the workflow of the solution includes six steps, namely: ① model migration domain determination, ② migration model selection, ③ application clustering, ④ model aggregation, ⑤ model callback, and ⑥ model migration.

[0026] The core modules of this solution are the design of multi-application traffic prediction model, application clustering algorithm and edge model migration strategy. The three core modules are described as follows:

[0027] 1. Multi-application traffic prediction model: This model is the basic support module of the embodiment of the present invention. All functions rely on this model. Its detailed architecture diagram is as follows: Figure 3 The model receives historical traffic sequence data of multiple applications as input and simultaneously outputs the traffic requirements of multiple applications at multiple time steps in the future, which can meet the fine-grained traffic management requirements at the application level.

[0028] Specifically, the model is generally designed based on the Transformer deep neural network architecture, which includes two components: encoder and decoder. The encoder is responsible for extracting features from historical input sequence data, encoding the learned high-dimensional abstract knowledge and passing it to the decoder. The decoder integrates the encoder's output with its own input, and then passes it to the decoder layer for learning and result output.

[0029] Both the encoder and decoder are composed of multiple stacked layers, among which the most important model designs include:

[0030] Multi-head attention mechanism: This design consists of multiple dot-product attention mechanisms to mine long-sequence temporal dependencies in historical inputs. The multi-head attention mechanism divides the data in its operation into blocks, each of which can be processed in parallel to speed up the training process, and finally merges the output results of each sub-block to pass the results to the next dot-product feedforward network layer.

[0031] Position encoding: The Transformer architecture design does not include a sequence structure design similar to a recurrent neural network. Therefore, position encoding needs to be introduced to implement the relative position marking of historical input sequence data, so as to introduce time series information into the prediction model.

[0032] Parallel encoding and masking: The traditional encoder-decoder architecture design uses the traditional Auto-Regressive concept to realize the encoding relationship of each time step of the time series, and does not have the ability to process in parallel. This model design adopts a multi-head attention mechanism with feedforward masking, which completes parallel encoding and training by masking the decoder input with historical time steps. Specifically, the decoder input is a sliding window of the final prediction sequence with a right shift of 1 time step. Feedforward masking prevents information leakage by setting future time series information to zero. That is, feedforward masking ensures that the encoder can only see the sequence of the previous t-1 time steps at time t, and all the sequences after time step t are set to zero, [y0, y1…, y t ,0,0,0…].

[0033] 2. Application clustering algorithm: Application clustering algorithm is used to cluster multiple applications in the same area according to the similarity of their time series. Multiple applications in each cluster are used to train one of the above-mentioned traffic prediction models to enhance the prediction results and alleviate the underfitting problem of the model.

[0034] In a system with large information entropy, data distribution is more chaotic. Under the premise of certain model representation ability, it is not possible to learn enough patterns so that the model is underfitting. Classifying the system applications according to the similarity of their traffic sequences can reduce information entropy, thereby alleviating the above problems. Specifically, the embodiment of the present invention designs the following application clustering algorithm:

[0035]

[0036] 3. Edge model migration strategy: This strategy aims to improve the generalization performance and reusability of the model, thereby reducing the operator's load overhead during data collection, processing and model training.

[0037] Specifically, this strategy mainly covers two steps: migration domain determination and migration model selection.

[0038] Migration domain determination: The migration of the model must be completed within the domain. The domain is determined based on the similarity of application distribution between edge nodes. The maximum, minimum, mean, and standard deviation of traffic in the region over a period of time are used as four features, and the K-means clustering algorithm is input to complete the domain segmentation of edge nodes. The edge nodes in each domain have similar statistical characteristics.

[0039] Migration model selection: Each edge node in the domain can use its own local data to maintain the traffic prediction model in the region, but this approach will cause great overhead. The model of the region with the largest amount of traffic is selected as the reference migration model among similar edge nodes in the domain, and it is migrated to other nodes in the domain for direct deployment. The richer the data in the region with large traffic, the more time series patterns the model can learn, so it has stronger universal performance. The experimental results verify this conclusion.

[0040] An application-level traffic prediction and model migration method for network edge provided by an embodiment of the present invention can be deployed on an edge computing node to predict multi-application network traffic in a region, and at the same time, the model migration deployment in the migration domain is completed through a model migration strategy, thereby completing fine-grained, application-level network traffic prediction on the basis of extremely low overhead to serve and dynamically allocate network resources, optimize network applications, and develop green networks.

[0041] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following will be described based on the specific implementation method of the present invention.

[0042] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of the present invention and its equivalent technology, the present invention is also intended to include these changes and variations.

Claims

1. A method for application-level traffic prediction and model migration at the edge of a network, characterized in that: The following steps are involved: Determine the model migration domain and use an unsupervised clustering algorithm based on application traffic statistical characteristics to complete the clustering of edge nodes; Select a migration model, where the migration model is a universal model selected in the migration domain to be migrated to each edge computing node; The traffic sequences of various network applications under each edge computing node are classified based on their similarity in time and shape distribution according to the application clustering algorithm; After classifying multiple network applications under each edge node, N categories are obtained. Each category trains its own multi-application traffic prediction model, and finally generates N multi-application traffic prediction models. These models are saved and wait for edge nodes to call them. When the training of N multi-application traffic prediction models is completed, they are called back by the edge computing node; In the migration domain, N multi-application traffic prediction models are migrated to all nodes in multiple domains for deployment and application.

2. The method according to claim 1, characterized in that According to the application clustering algorithm, multiple applications in the same area are clustered according to the similarity of their time series, and multiple applications in each cluster are used to train the above-mentioned traffic prediction model to enhance the prediction results and alleviate the under-fitting problem of the model.

3. The method according to claim 1, characterized in that The multi-application traffic prediction model includes two components: an encoder and a decoder. The encoder is responsible for extracting features from historical input sequence data, encoding the learned high-dimensional abstract knowledge and passing it to the decoder. The decoder integrates the encoder's output with its own input and then passes it to the decoder layer for learning and result output.

4. The method according to claim 3, characterized in that The encoder and the decoder are respectively composed of a plurality of stacked layer structures, including: Multi-head attention mechanism: It is composed of multiple dot product attention mechanisms to mine long-sequence temporal dependencies in historical inputs. The multi-head attention mechanism divides the data in its operation process into blocks, and each data block can be processed in parallel to accelerate the training process. Finally, the output results of each sub-block are merged to pass the results to the lower dot product feedforward network layer. Position encoding: realizes the relative position marking of historical input sequence data, thereby introducing time series information into the prediction model; Parallel encoding and masking: A multi-head attention mechanism with feedforward masking is used to complete parallel encoding and training by masking the decoder input with historical time steps; the decoder input is a sliding window of the final prediction sequence with a right shift of 1 time step; feedforward masking prevents information leakage by setting future time series information to zero.

5. The method according to claim 1, characterized in that The edge model migration strategy covers two steps: migration domain determination and migration model selection: Migration domain determination: The migration of the model must be completed within the domain. The domain is determined based on the similarity of application distribution between edge nodes. The maximum, minimum, mean, and standard deviation of traffic in the region over a period of time are used as four features, and the K-means clustering algorithm is input to complete the domain segmentation of edge nodes. The edge nodes in each domain have similar statistical features. Migration model selection: Each edge node in the domain can use its own local data to maintain the traffic prediction model for the region, but this approach will cause great overhead. The model of the region with the largest amount of traffic is selected as the reference migration model among similar edge nodes in the domain, and then migrated to other nodes in the domain to directly deploy the application.

6. The method according to claim 1, characterized in that The step of clustering edge nodes using an unsupervised clustering algorithm based on application traffic statistical features includes: For each application, the maximum, minimum, mean, and standard deviation of the traffic sequence are used as four statistical features. For N applications, the dimension of the input matrix of the K-means algorithm is [4*N]; the clustering result determines the optimal number of clusters based on the silhouettescore, and its value range is [-1,1]; the edge nodes in each cluster show similarity in the statistical features of multi-application traffic.

7. The method according to claim 1, characterized in that The selected migration model generates a model trained by traffic from multiple applications in the domain and the largest node.

8. The method according to claim 1, characterized in that The step of classifying the similarities in time and shape distribution of traffic sequences of various network applications under each edge computing node according to the application clustering algorithm includes: The similarity between applications is determined based on the Wasserstein distance. Compared with the traditional Euclidean distance, this distance can take into account the similarity of the shape distribution between two sequences, thereby more accurately classifying applications to reduce the information entropy within the class, thereby alleviating the problem of underfitting of the model.

Images