Research On Traffic Allocation Algorithm Based On Machine Learning With Network Slicing

With the fast growth of wireless network technologies(e.g.,5G)and increasing demand for services with high Quality of Services(Qo S),efficient management of network resources becomes more and more important.Numerous studies have been done to improve Qo S and network resource utilization for traffic scheduling.A helpful way to achieve this target is queue management and scheduling of network traffic.However,the queuing nature of the queue management scheduling have an implicit drawback.As the feasible routes for flows need to be calculated real-time by checking the remaining bandwidth over network links,the time for computing a feasible route may be time-consuming as flows in the queue are sequentially processed.It has been noticed that network slicing have an effect of acceleration on routing.Observing this,a network slicing model is introduced in this paper to accelerate routing.In the proposed slicing model,different slices share the same network topology but the bandwidth resources can be different on the links.Flows are partitioned into different slices.As flows are processes independently in each slice,the computation of routes can be in parallel and this accelerates the routing process.The main challenge in implementing such an idea is that,when determining which slice a flow should be deployed in,it should guarantee a high success rate of transmission as some flows may fail to be transmitted due to resource scarcity.To address this problem,we propose a supervised-learning-based model to predict which slice a request should be deployed in,so as to maximize the number of successfully transmitted flows.Based on the technical implementation of machine learning,the proposed method predicts a batch of flows at a time,and then allocates the flows to the specified slices according to the predicted results.The main contributions of this paper are summarized as follows:(1)A slicing model that makes different slices share the same network topology is proposed,which can help us achieve the purpose of reducing the time it takes to calculate routes for flows by parallel computing.Based this slicing model,we establish the model of flows allocation in network slices with the aim of maximizing the number of successfully transmitted flows and total amount of data transmitted.(2)Aiming at the flows allocation problem studied in this paper,a neural network model is proposed.This model can decide which slice each flow should be deployed in according to the network slice state and information of multiple flows.The model includes a five-layer neural network,which has the characteristics of multi-output and partial parameter sharing of the neural network.In order to provide effective training samples for the model,a sample generation algorithm is also proposed in this paper.(3)In order to solve the problem that the effect of the above model becomes worse when network slices have different reamining bandwidth,a neural network model based edge graph convolution is proposed.By changing the node-based graph convolution into edge-based graph convolution,the model can extract the state information and spatial structure information on the network slice link.In order to further improve the performance,multiple edge-based graph convolution neural networks are stacked in the model and the jump connection is introduced.(4)In order to evaluate the effectiveness of the algorithm in this paper,experimental evaluations were conducted in a simulation environment based on a 5G network scenarios.Experimental results show that our approaches can achieve good performance on reducing time to calculate routes,improving the success rate of transmission and increasing total amount of data transmitted.