Master of Engineering Science
Electrical and Computer Engineering
With the growing demand for data connectivity, network service providers are faced with the task of reducing their capital and operational expenses while simultaneously improving network performance and addressing the increased connectivity demand. Although Network Function Virtualization has been identified as a potential solution, several challenges must be addressed to ensure its feasibility. The work presented in this thesis addresses the Virtual Network Function (VNF) placement problem through the development of a machine learning-based Delay-Aware Tree (DAT) which learns from the previous placement of VNF instances forming a Service Function Chain. The DAT is able to predict VNF instance placements with an average 34μs of additional delay when compared to the near-optimal BACON heuristic VNF placement algorithm. The DAT’s max depth hyperparameter is then optimized using Particle Swarm Optimization (PSO) and its performance is improved by an average of 44μs through the introduction of the Depth-Optimized Delay-Aware Tree (DO-DAT).
Summary for Lay Audience
The past two decades have seen an incredible increase in the number of devices producing network traffic including: smartphones, tablets, wearable smart devices, and smart home accessories. This surge in network traffic puts a major burden on service providers world-wide. In order to keep up with the demand, these providers must make significant upgrades on their network infrastructure. This proves to be a challenging task as there are several functionalities (e.g. firewalls) offered by the network which require infrastructure specific to those functions. Upgrading these specific infrastructural components and increasing the number of each present in the network is an extremely costly endeavour which results in significant capital expenditures incurred by the service providers. As a way to mitigate these costs as well as improve system performance and network strength, Network Function Virtualization (NFV) has been proposed as a solution. This technology essentially isolates the functionality of each of these specific infrastructural components and turns them into software applications which can be run on generic infrastructure such as datacenter servers. There are several issues which arise from this technology which must be addressed to ensure its feasibility. One of these includes the placement of these software applications in the network. While traditionally this placement task has been achieved through optimization models and approximate solutions, these can often require significant time and resources to solve. This is inadequate for network systems as there are several time critical applications using the network. As an alternative, this work outlines the use of machine learning to make a model which predicts the placement of each of these software applications based on previous placements. Quantitative results show that the machine learning model can predict a placement which produces a slightly higher delay between applications compared to a current approximate solution. To mitigate this, a domain based optimization model is presented to optimize the parameters of the previous machine learning model such that the placement results in reduced delay between applications. Results show a significant improvement once optimized and confirm that the work presented is a significant step towards a fully automated placement strategy.
Manias, Dimitrios Michael, "Machine Learning for Performance Aware Virtual Network Function Placement" (2019). Electronic Thesis and Dissertation Repository. 6482.