MUNA AL-SAADI

Abstract

The main task of a network is to hold and transfer data between its nodes. To achieve this task, the network needs to find the optimal route for data to travel by employing a particular routing system. This system has a specific job that examines each possible path for data and chooses the suitable one and transmit the data packets where it needs to go as fast as possible. In addition, it contributes to enhance the performance of network as optimal routing algorithm helps to run network efficiently. The clear performance advantage that provides by routing procedures is the faster data access. For example, the routing algorithm take a decision that determine the best route based on the location where the data is stored and the destination device that is asking for it. On the other hand, a network can handle many types of traffic simultaneously, but it cannot exceed the bandwidth allowed as the maximum data rate that the network can transmit. However, the overloading problem are real and still exist. To avoid this problem, the network chooses the route based on the available bandwidth space. One serious problem in the network is network link congestion and disparate load caused by elephant flows. Through forwarding elephant flows, network links will be congested with data packets causing transmission collision, congestion network, and delay in transmission. Consequently, there is not enough bandwidth for mice flows, which causes the problem of transmission delay. Traffic engineering (TE) is a network application that concerns with measuring and managing network traffic and designing feasible routing mechanisms to guide the traffic of the network for improving the utilization of network resources. The main function of traffic engineering is finding an obvious route to achieve the bandwidth requirements of the network consequently optimizing the network performance [1]. Routing optimization has a key role in traffic engineering by finding efficient routes to achieve the desired performance of the network [2]. Furthermore, routing optimization can be considered as one of the primary goals in the field of networks. In particular, this goal is directly related to traffic engineering, as it is based on one particular idea: to achieve that traffic is routed according to accurate traffic requirements [3]. Therefore, we can say that traffic engineering is one of the applications of multiple improvements to routing; routing can also be optimized based on other factors (not just on traffic requirements). In addition, these traffic requirements are variable depending on analyzed dataset that considered if it is data or traffic control. In this regard, the logical central view of the Software Defined Network (SDN) controller facilitates many aspects compared to traditional routing. The main challenge in all network types is performance optimization, but the situation is different in SDN because the technique is changed from distributed approach to a centralized one. The characteristics of SDN such as centralized control and programmability make the possibility of performing not only routing in traditional distributed manner but also routing in centralized manner. The first advantage of centralized routing using SDN is the existence of a path to exchange information between the controller and infrastructure devices. Consequently, the controller has the information for the entire network, flexible routing can be achieved. The second advantage is related to dynamical control of routing due to the capability of each device to change its configuration based on the controller commands [4]. This thesis begins with a wide review of the importance of network performance analysis and its role for understanding network behavior, and how it contributes to improve the performance of the network. Furthermore, it clarifies the existing solutions of network performance optimization using machine learning (ML) techniques in traditional networks and SDN environment. In addition, it highlights recent and ongoing studies of the problem of unfair use of network resources by a particular flow (elephant flow) and the possible solutions to solve this problem. Existing solutions are predominantly, flow routing-based and do not consider the relationship between network performance analysis and flow characterization and how to take advantage of it to optimize flow routing by finding the convenient path for each type of flow. Therefore, attention is given to find a method that may describe the flow based on network performance analysis and how to utilize this method for managing network performance efficiently and find the possible integration for the traffic controlling in SDN. To this purpose, characteristics of network flows is identified as a mechanism which may give insight into the diversity in flow features based on performance metrics and provide the possibility of traffic engineering enhancement using SDN environment. Two different feature sets with respect to network performance metrics are employed to characterize network traffic. Applying unsupervised machine learning techniques including Principal Component Analysis (PCA) and k-means cluster analysis to derive a traffic performance-based clustering model. Afterward, thresholding-based flow identification paradigm has been built using pre-defined parameters and thresholds. Finally, the resulting data clusters are integrated within a unified SDN architectural solution, which improves network management by finding the best flow routing based on the type of flow, to be evaluated against a number of traffic data sources and different performance experiments. The validation process of the novel framework performance has been done by making a performance comparison between SDN-Ryu controller and the proposed SDN-external application based on three factors: throughput, bandwidth,and data transfer rate by conducting two experiments. Furthermore, the proposed method has been validated by using different Data Centre Network (DCN) topologies to demonstrate the effectiveness of the network traffic management solution. The overall validation metrics shows real gains, the results show that 70% of the time, it has high performance with different flows. The proposed routing SDN traffic-engineering paradigm for a particular flow therefore, dynamically provisions network resources among different flow types.

Keywords

Unsupervised Machine Learning (ML), Traffic analysis, Software Defined Networks (SDN), Data Center Networks (DCN), K-means, Principal Components Analysis (PCA), Elephant flows, Mice flows, Flows identification, SDN-application, Network Performance Analysis (NPA)

Document Type

Thesis

Publication Date

2023

Download

Share

COinS