A data-driven method for future Internet route decision modeling

https://doi.org/10.1016/j.future.2018.12.054Get rights and content

Highlights

  • We propose a data-driven model for the inter-domain route decision process with deep learning method, which learns, understands, and models the route decision process without the priori knowledge.

  • We propose a set of deep learning resolution to our model with structure, characterization, feature selection, and training data construction.

  • We discuss the effectiveness of our paper with detailed cases, indicating that our model outperforms the AS business relationship model.

Abstract

Simulating the BGP routing system of Internet is crucial to the analysis of Internet backbone network routing behavior, locating network failure and, evaluating network performance for future Internet. However, the existing BGP routing model lacks in the coarse modeling granularity and the priori knowledge based model. The analysis of BGP routing data that reflects the routing behaviors, directly impacts the BGP routing decision and forward strategy. The efficiency of such analysis dictates the time it takes to come up with such a time-critical decision and strategy. Under the existing model, BGP routing data analysis does not scale up.

In this paper, we analyze the inter-domain routing decision making process, then present a prefix level route decision prediction model. More specifically, we apply deep learning methods to build a high-precision BGP route decision process model. Our model handles as much available routing data as possible to promote the prediction accuracy. It analyzes the routing behaviors without any prior knowledge. Beyond discussing the characteristics of the model, we also evaluate the proposed model using experiments explained in detailed cases. For the research community, our method could help in detecting routing dynamics and route anomalies for routing behavior analysis.

Introduction

Internet is composed of tens of thousands of Autonomous Systems (AS). Such ASes run their network individually, and exchange their routing information using inter-domain routing protocol, which is Border Gateway Protocol (BGP) in real-life deployment. In spite of the growing trend of discussing the next generation networking, the distributed management framework of Internet can hardly be changed due to its economic, political, geographical involvements.

Inter-domain routing protocol plays a dominant role in the maintenance and management of Internet. Appropriate protocol configurations could significantly improve network performance. On the contrary, inappropriate protocol configuration could be a disaster to the regional network, or even the whole Internet. It has been decades since the research community realized the importance of understanding, analyzing, and predicting the inter-domain routing behaviors, then modeling BGP network’s route decision process. Unfortunately, modeling the route decision process is a non-trivial problem. An AS’s BGP configuration involves its business secret, thus AS administrators never share their network’s BGP configuration. The only way to conduct route decision modeling is to compare the input and output of the route decision process, then construct a general mapping from the input to the output. However, for BGP, the same route decision result could be reached after a series of equivalent configurations on the granularity of prefix. Due to different economic interest, traffic engineering objectives and political reasons, practical configurations of BGP are always different. Therefore, to form a general routing model for BGP’s route decision process that satisfies most cases is always a challenge.

Generally, existing work models the route decision process based on prior experience, which means analyzing data, making assumptions, building a model, and finally verifying the correctness of the model. Such manner takes the route decision process as a white box, which explains explicitly how and why the model works. However, it leads to the following limitations of the route decision model:

First, researchers have to make the right assumption in the first place. Due to the reasons discussed above, such assumption can hardly fit all ASes de facto BGP configurations, thus limits the accuracy of the model. For example, AS business relationship model, the most famous routing model, assumes that the AS administrators always run their network based on their business relationships with their network’s neighboring ASes. The model has been proven to be correct. However, recent work has found more and more counterexamples of this AS business relationship model. Growing amount of supporting BGP data has indicated that the de facto AS relationship is actually more complex than the AS business relationship. As a result, the AS business relationship model does not work for applications that need understanding of certain routing behaviors in a finer granularity (e.g. in prefix level).

Second, since adjustment on the model according to the input/output data is an inevitable , the structure of the white box (the model) must not be too complex to conduct any adjustment. As a result, the estimation of the model parameters can only be performed on a limited quantity of BGP data. Thus the accuracy of the route decision model can hardly be improved when the available BGP data increases. For example, since 2001, when the AS business routing model was first proposed, the amount of observable BGP data has been increased by 2 orders of magnitude, and the amount of BGP data is still growing. Such available data should reveal more detailed information of an inter-domain routing system. However, due to the limited expressiveness of the existing white box routing models, the accuracy and performance of the existing routing models are limited.

In recent years, deep learning technology is developing rapidly, which gives us an opportunity to model the BGP route decision process in a smarter way, i.e. to conduct data-driven modeling on the route decision process directly from the routing data, without understanding or explaining everything. With deep learning methods, we can form a general classification model to work as the route decision process. It learns from the available BGP data all by itself, then reveals possible configurations of BGP protocol. Intuitively, the accuracy of the route decision process model should be improved with the growing amount of BGP data, since theoretically, the more input data fed to the neural network structure, the more accurate the results. However, few existing works dedicate to model the route decision process by means of deep learning methods.

Considering the limitations of existing white box routing models, we propose to view the route decision process as a black box, and try to solve the route decision process modeling problem using deep learning methods. Since the route decision process takes the candidate routes as inputs, then outputs the optimal route, we model it as a classifier which distinguishes the optimal route from other candidate routes. Fundamentally, the challenges of this modeling problem include (1) model structure rationality discussion and, (2) parameter estimation for the model for each AS. In this paper, we focus on the former. We also discuss the model structure, the characterization of candidate routes, the training data set construction as well as the model evaluation.

The contributions of this paper are three folds:

(1) We propose to model the route decision process using a data-driven method, which enables us to focus on the efficiency of our networking model, without the explanation of model structure during the construction phase.

(2) We propose an efficient supervised learning resolution for the route decision process modeling, including the characterization, feature selection, and training data set construction modules. Our deep learning resolution ensures the scalability of the route decision process modeling, so that the model accuracy improves with the growth of the available BGP data.

(3) We investigate the feasibility of our model with open source BGP data based case study evaluation. We also compare our model with the AS business relationship model, proving its effectiveness for route decision modeling in finer granularity. We then discuss possible further applications of the proposed model in data analysis, network modeling and prediction.

The rest of the paper is organized as following. First we introduce related works in Section 2. Then we propose the general structure and details of our routing model, including the characterization, the training set construction, and how to tag the training data in Section 3. Next, we evaluate our model with comparison to the AS business relationship model using a case study presented in Section 4. Finally, we discuss the feasibility of our model in Section 5, then conclude the paper with future works to follow in Section 6.

Section snippets

Route decision process modeling

Researchers have been pursuing an appropriate model for inter-domain routing policy during the last two decades. Lixin Gao proposed AS business relationship [1], which is the first and most widely-used model on inter-domain routing policy. She partitioned routing policy into 3 classes: provider–customer, peer–peer, and sibling–sibling. Based on AS business relationship, she proved that a reasonable AS path should follow valley-free policy. Based on this model, there have been plenty of works on

Design of our route decision model

BGP’s route decision process learns candidate routes from neighboring routers, ranking the preference of the candidate routes, and decides the optimal route. Herein, as a general model for the route decision process, our model takes the candidate routes information as input, and decides the optimal route as output.

Since BGP’s route decision process decides the optimal for each prefix, our model also works in a per-prefix level. Since the route decision process ranks all candidate routes

Evaluation by case study

In this section, we investigate the feasibility of our proposed model. As a first step to the correctness of our method, we want to begin with case study of routing policy modeling, focusing on the detailed routing scenarios, and investigate the difference between our method and the previous proposed methods. To that end, we first introduce our used data set together with the topology of our example cases; we then compare the prediction accuracy of our method with the AS business relationship

Model characteristic

The AS business relationship model make the route choice according to the next hop AS type, and the AS path length. We believe such two factors are both considered in our model. The neighboring ASes are the most frequently observed ASes in monitor’s routing table, so they have first priority to be selected as features. With further appropriate training, the routes from the preferred neighboring ASes are more likely to be selected by our model since it simply speaks for the data.

Since our model

Conclusion

In this paper, we introduce a data-driven method for the BGP network modeling, which sheds light on modeling the BGP route decision process as a black box. We discuss the modeling details including model structure, route set construction, characterization and feature selection. By comparison with the AS business relationship model in the form of a case study, we prove the effectiveness of our model. As a future work, it is necessary to conduct evaluation in more cases with more training data in

Acknowledgments

This work is supported by the National Natural Science Foundation of China under Grant No. 61572153, No. 61702220, and No. 61702223 and the National Key research and Development Plan (Grant No. 2018YFB0803504).

Zhihong Tian Ph.D., professor, PHD supervisor, Dean of cyberspace institute of advanced technology, Guangzhou University. Standing director of CyberSecurity Association of China. Member of China Computer Federation. From 2003 to 2016, he worked at Harbin Institute of Technology. His current research interest is computer network and network security. E-mail: tianzhihong@ gzhu.edu.cn.

References (30)

  • DuX. et al.

    An effective key management scheme for heterogeneous sensor networks

    Ad Hoc Netw.

    (2007)
  • KohaviR. et al.

    Wrappers for feature subset selection

    Artificial Intelligence

    (1997)
  • L. Gao, On inferring Autonomous System relationships in the Internet, in: IEEE/ACM Transactions on Networking,...
  • L. Subramanian, S. Agarwal, J. Rexford, R.H. Katz, Characterizing the Internet hierarchy from multiple vantage points,...
  • G.D. Battista, M. Patrignani, M. Pizzonia, Computing the types of the relationships between Autonomous Systems, in:...
  • T. Erlebach, A. Hall, T. Schank, Classifying customer-provider relationships in the Internet, in: Proc. of the IASTED...
  • G. Siganos, M. Faloutsos, Analyzing BGP policies: Methodology and tool, in: IEEE INFOCOM,...
  • Z. Morley Mao, L. Qiu, J. Wang, Y. Zhang, On AS-level path inference, in: Proc. ACM SIGMETRICS,...
  • Y. Afek, O. Ben-Shalom, A. Bremler-Barr, On the Structure and application of BGP Policy Atoms, in: IMW02: Proc. ACM...
  • A. Broido, K.C. Claffy, Analysis of RouteViews BGP Data: Policy Atoms, in: Proc. of the Network-Related Data Management...
  • W. Mubauauer, S. Uhlig, B. Fu, M. Meulle, O. Maennel, In search for an appropriate granularity to model routing...
  • GuptaA. et al.

    Sdx: a software defined internet exchange

    ACM SIGCOMM Comput. Commun. Rev.

    (2015)
  • MazloumR. et al.

    Violation of interdomain routing assumptions

  • AnwarR. et al.

    Investigating interdomain routing policies in the wild

  • C. Cardona, P. Francois, P. Lucente, Impact of BGP filtering on Inter-Domain Routing Policies (No. RFC 7789),...
  • Cited by (0)

    Zhihong Tian Ph.D., professor, PHD supervisor, Dean of cyberspace institute of advanced technology, Guangzhou University. Standing director of CyberSecurity Association of China. Member of China Computer Federation. From 2003 to 2016, he worked at Harbin Institute of Technology. His current research interest is computer network and network security. E-mail: tianzhihong@ gzhu.edu.cn.

    Shen Su born in 1985, Ph.D., assistant professor, Guangzhou University. His current research interests include inter-domain routing and security, Internet of connected vehicles, and wireless sensor networks. E-mail: [email protected].

    Wei Shi is an assistant professor at the University of Ontario Institute of Technology. She is also an adjunct professor at Carleton University. She holds a Bachelor of Computer Engineering from Harbin Institute of Technology in China and received her Ph.D. of Computer Science degree from Carleton University in Ottawa, Canada.

    Xiaojiang Du Dept. of Computer and Information Sciences, Temple University, Philadelphia PA 19122, USA, email: [email protected]. His research interests are wireless communications, wireless networks, security, and systems.

    Mohsen Guizani Dept. of Electrical and Computer Engineering, University of Idaho, Moscow, Idaho, USA, E-mail:[email protected]. His research interests include wireless communications and mobile computing, computer networks, mobile cloud computing, security, and grid.

    Xiang Yu born in 1978, Ph.D., Heilongjiang Institute of Technology. His current research interest is network and information security. E-mail: [email protected].

    View full text