Evaluation of mathematical models to estimate end‐to‐end traffic in a backbone network

Abstract

End-to-end network traffic analysis is one of the most important factors in design, development, and traffic engineering of large-scale backbone networks. Despite the importance of end-to-end traffic, it is not applicable to measure it directly using protocols such as NetFlow in large-scale networks due to high computational cost. One technique to tackle this problem is estimating endto-end traffic using easily measured neighboring link loads. Tomography model relates these two parameters with linear equations. In this way, the problem turns into an inverse problem. Since the number of link loads is usually smaller than the number of end-to-end flows in a computer network, traffic matrix estimation is considered to be an ill-posed problem. In this paper, we propose two different methods for estimation of end-to-end traffic matrix by means of link load measurements. One of the methods deploys Kalman filter and the other is developed based on Tikhonov regularization with respect to the existence of temporal autocorrelation among traffics. We also propose a best estimation model based on least squares, which explores the temporal autocorrelation in end-to-end traffics. This model is used as the best estimator and has been compared with other methods. We have evaluated the efficiency of the proposed methods using real datasets from Abilene and GÉANT backbone networks. It is worth noting that the proposed methods in this paper also can be used in any inverse problem with temporal autocorrelation

Spatiotemporal traffic matrix prediction: A deep learning approach with wavelet multiscale analysis

“…In addition, we further evaluate the relative prediction error of each OD flow during the time individually, and the relative prediction error of all OD flows at each time‐slot. We use two popular error evaluation metrics, namely, Spatial relative error (SRE) and temporal relative error (TRE). Spatial relative error denotes each OD flows relative error during the whole observation time‐slots, and TRE represents how prediction errors evolve over time.…”

“…NetFlow/sFlow requires proprietary software configuration or hardware installation, which inevitably has a large impact on the performance of forwarding devices . Even in the emerging software‐defined network (SDN), most existing traffic monitoring techniques achieve network status monitoring through extra probe packets or proprietary protocols, making the network monitoring cost too much . Therefore, we cannot obtain network‐wide traffic easily and directly.…”

“…To achieve accurate network‐wide traffic forecasting, also called TM prediction, network operators need to model the spatiotemporal features (dependencies) within TM series well, namely, the elements of a TM are usually high correlated in space (ie, spatial features) and their variability over time may be characterized by complex nonlinear evolution patterns at different time scales (ie, temporal features) . Note that we do not address the TM estimation problem in this paper, which aims to solve an inverse inference problem with the highly ill‐posed property, and we suppose that historical TM data is already obtained.…”

A new approach for traffic matrix estimation in high load computer networks based on graph embedding and convolutional neural network

“…Vardi introduced tomography model, in which he spotted a linear relation between the link loads and OD flows via a routing matrix. Almost all of the traditional end‐to‐end traffic estimators were based on this linear model . Using tomography model, one can accurately calculate the link loads by having the OD flows and the routing matrix in hand.…”