| Nowadays,diversified service applications put forward strict requirements on network performance.Among all kinds of network performance indicators,the most concerned one is network delay,which can accurately reflect the network service quality level.In small-scale network system,pure delay measurement method is used to obtain delay data,but the measurement result is affected by asymmetric one-way delay and the error is large.In large scale network system,time delay data is usually obtained by "less measurement,more prediction" method,and feature selection is closely related to algorithm learning and prediction ability.Based on the above situation,this paper studies one-way delay measurement in asymmetric network environment and delay prediction method in large-scale network.The main work of this paper is as follows:(1)Clock offset is the biggest obstacle to time delay measurement,especially in an asymmetric network environment,it is difficult to obtain the clock offset between two hosts by simply processing the timestamp information of the measurement packet.If the asymmetry is mainly caused by the randomness of queuing delay,this paper proposes a one-way delay measurement algorithm based on the elimination of queuing delay.The algorithm calculates queuing delay according to the minimum delay,so as to remove the influence of network asymmetry on the delay measurement algorithm and improve the measurement accuracy.At the same time,a clock reset point detection algorithm based on piecewise linear representation is proposed.According to the detection results,the time delay sequence is divided into several homogeneous segments and the clock skew in the segments is solved respectively,so as to obtain high-precision clock skew calculation results.(2)The delay prediction method based on some machine learning only relies on experience to select artificial features,which has problems such as subjectivity,inability to determine whether the selection is optimal,and the investment in manpower and time is also a key factor that cannot be ignored.In view of the above problem,this paper proposes a delay prediction method based on fusion(GBDT-SVR)model,The mayfly algorithm is used to optimize the parameters of the support vector regression model.At the same time,combined with the significant advantages of gradient boosting decision tree in feature processing and the strong learning ability and generalization ability of support vector regression model,the problem of poor accuracy of artificial time delay feature selection is effectively made up.(3)The clock reset detection algorithm based on piecewise linear representation,one-way delay measurement algorithm based on queuing delay elimination and delay prediction algorithm based on fusion model proposed in this paper are verified and analyzed respectively.The experimental results show that the clock reset detection algorithm proposed in this paper can effectively avoid the missing and false detection of reset points by existing algorithms;The proposed delay measurement algorithm has less error than the traditional method,and the proposed fusion model has improved prediction accuracy compared with the single model,which has important significance and research value. |
