Knowledge-defined networking: a machine learning based approach for network and traffic modeling
- Autores: Albert Mestres Sugrañes
- Directores de la Tesis: Albert Cabellos Aparicio (dir. tes.), Eduardo Alarcón (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Wouter Tavernier (presid.), Josep Solé Pareta (secret.), José Núñez Martínez (voc.)
- Programa de doctorado: Programa de Doctorado en Arquitectura de Computadores por la Universidad Politécnica de Catalunya
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
The research community has considered in the past the application of Machine Learning (ML) techniques to control and operate networks. A notable example is the Knowledge Plane proposed by D.Clark et al. However, such techniques have not been extensively prototyped or deployed in the field yet. In this thesis, we explore the reasons for the lack of adoption and posit that the rise of two recent paradigms: Software-Defined Networking (SDN) and Network Analytics (NA), will facilitate the adoption of ML techniques in the context of network operation and control. We describe a new paradigm that accommodates and exploits SDN, NA and ML, and provide use-cases that illustrate its applicability and benefits. We also present some relevant use-cases, in which ML tools can be useful. We refer to this new paradigm as Knowledge-Defined Networking (KDN).
In this context, ML can be used as a network modeling technique to build models that estimate the network performance. Network modeling is a central technique to many networking functions, for instance in the field of optimization. One of the objective of this thesis is to provide an answer to the following question: \emph{Can neural networks accurately model the performance of a computer network as a function of the input traffic?}. In this thesis, we focus mainly on modeling the average delay, but also on estimating the jitter and the packets lost. For this, we assume the network as a black-box that has as input a traffic matrix and as output the desired performance matrix. Then we train different regressors, including deep neural networks, and evaluate its accuracy under different fundamental network characteristics: topology, size, traffic intensity and routing. Moreover, we also study the impact of having multiple traffic flows between each pair of nodes.
We also explore the use of ML techniques in other network related fields. One relevant application is traffic forecasting. Accurate forecasting enables scaling up or down the resources to efficiently accommodate the load of traffic. Such models are typically based on traditional time series ARMA or ARIMA models. We propose a new methodology that results from the combination of an ARIMA model with an ANN. The Neural Network greatly improves the ARIMA estimation by modeling complex and nonlinear dependencies, particularly for outliers. In order to train the Neural Network and to improve the outliers estimation, we use external information: weather, events, holidays, etc. The main hypothesis is that network traffic depends on the behavior of the end-users, which in turn depend on external factors. We evaluate the accuracy of our methodology using real-world data from an egress Internet link of a campus network. The analysis shows that the model works remarkably well, outperforming standard ARIMA models.
Another relevant application is in the Network Function Virtualization (NFV). The NFV paradigm makes networks more flexible by using Virtual Network Functions (VNF) instead of dedicated hardware. The main advantage is the flexibility offered by these virtual elements. However, the use of virtual nodes increases the difficulty of modeling such networks. This problem may be addressed by the use of ML techniques, to model or to control such networks. As a first step, we focus on the modeling of the performance of single VNFs as a function of the input traffic. In this thesis, we demonstrate that the CPU consumption of a VNF can be estimated only as a function of the input traffic characteristics.
