Proactive network and traffic aware network optimization

Abstract

The Information and Communication Technology (ICT) sector is rapidly becoming a big contributor among the energy-consuming sectors. Recent studies have shown that the ICT impact on global power consumption is non-negligible - about 4% (GeSi SMARTer2030, n.d.). This is due to a range of new technologies, including moving from 4G to 5G, the advent of the Internet of Things (IoT), the proliferation of electronic devices and the cost-effective production processes needed to manufacture such devices. Telecommunication networks consume 37% of ICT power consumption. In many developed countries, ICT-related sectors are on the top list of power consumers. Moreover, an increase in global ICT-related electricity consumption is projected by 2030 compared to 2020. Traditionally, telecommunication networks have been designed to maximize the available bandwidth. This policy tends to minimize the replacement costs that occur when technologies are updated to increase either in the number of users or in the exchanged traffic. However, users access the network at different times of the day, and the applications usages are different in terms of traffic loads. Networks are dimensioned for the highest traffic demands so that the quantity of data flowing in the network usually is guaranteed below the maximum achievable data rates. The power consumption of network devices depends on the installed capacity of the underlying technology. The key idea of the design and operation of Energy-Aware (EA) networks is to reduce the difference between the network’s utilization and the provided capacity. In order to minimize energy waste, or equivalently, to make the consumption of the network equal to the traffic load, different solutions are being studied. The existing approaches include two main categories: 1. Energy-proportional approaches try to achieve energy proportionality by adapting the devices’ speed (and capacity) to the actual load. In this case, link speed is decreased when the link is underutilized because the interface cards with lower transmission rates consume smaller amounts of power. 2. Sleep mode approaches affect the network as a whole and approximate load proportionality by carefully distributing the traffic in the network so that some devices are fully utilized, and other devices become idle and switched into sleep mode. In this case, the current devices’ energy consumption is practically independent of the load. Therefore, switching off the devices can save a consistent amount of energy. However, sleep modes introduce additional complexity to the network because it needs coordination among devices. A famous method in energy-proportional approaches is known as Adaptive Link Rate (ALR). Adaptive link rate (ALR) is an effective means to save energy consumption of network elements by adjusting the link rate according to the carried traffic through a network-level optimization of the flow allocation process. In this research, we focus on ALR to provide a solution for the network energy efficiency problem. It is worth noting that current adaptation approaches are mainly reactive, in which link speed is adjusted when the traffic demand is changed. As the traffic flows fluctuate, the globally optimum network configuration changes over time, and the network requires reconfiguring to maintain the minimum power consumption. These approaches require multiple re-optimizations in each iteration which hurt Quality of Service (QoS) and network stability. In this thesis, we propose a framework based on the Rate Adaptation using Prediction (RAP) to predictively optimize link rates based on the multiple-step-ahead forecasting traffic demand. RAP finds a network configuration that minimizes the energy cost in the current time slot as well as requires the minimum modifications to be adjusted to the demands in future time slots. We formulate RAP as an integer linear programming (ILP) model and propose a heuristic simulated annealing (SA) algorithm to solve it. The model has been evaluated over two well-known network topologies (GEANT and fat-tree) using real-life traffic data. Our experimental results show that our approach provides a significant energy-saving while it decreases (on average) %61 of re-optimizations in the energy-aware routing. In this thesis, there are four notable contributions to this problem. First, we design a Gaussian Process Regression (GPR) kernel called semi-periodic self-similar (SPSS) covariance function, which is based on the traffic characteristics and shows a significant improvement in the traffic prediction accuracy compared to other methods. Second, we propose a GPR ensemble model to capture different volatile traffic patterns. The core of our proposed model is a novel method to optimize the accuracy-diversity balance between the base learners, which results in enhanced prediction performance. The third contribution is a multi-step-ahead prediction algorithm using the traffic characteristics. We showed how traffic’s multi-scale nature could be reflected in the multi-step-ahead prediction to reduce the error propagation in an extended prediction horizon. This approach outperforms existing powerful time-series algorithms such as LSTM. The fourth contribution is a network energy optimization model. While exiting energy efficiency optimization models consider only power consumption, we used traffic prediction to control and minimize the number of changes in the network during our online optimization solution. These four contributions are integrated into our proposed network energy framework, and we build a traffic-aware solution to reduce network power consumption with dynamic traffic.

Date	18 Aug 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Mohamed Cheriet (Supervisor)