An artificial intelligence powered framework for automatic service function chain placement in distributed scenarios

Abstract

Software Defined Network (SDN) and Network Function Virtualisation (NFV) are making net- works programmable and consequently much more flexible and agile. To meet service level agreements, achieve greater utilisation of legacy networks, faster service deployment, and re- duce expenditure, telecommunications operators are deploying increasingly complex Service Function Chains (SFCs). Besides the advantages from service virtualisation, it is expected that network performance and availability do not be affected by SFC usage. However, several factors that may compromise the SFC availability are added in a virtualised scenario such as software failures, misconfiguration, cyberattacks, and so on. In order to mitigate the impact of these factors, redundancy mechanisms can be used, i.e., to add redundant Virtual Network Functions (VNFs) in the servers to keep the SFC operation in case of failures. On the other hand, the network operators desire, of course, to allocate the SFCs optimising the resources utilisation in order to reduce Operational Expenditures (OPEX), which is a challenge since the replication mechanisms demand additional computational resources. In addition, the place- ment of SFCs in distributed scenarios can improve their availability, since an isolated failure would not impact the whole SFC operation. However, the placement in geo-distributed scenar- ios increases the management complexity, where different hardware and additional delay may compromise the SFC performance. Therefore, intelligent strategies are needed to optimise the SFC placement. This thesis presents the Sfc Placement framework focused on avaIlability for DistributEd scenaRios (SPIDER), a framework for SFC placement with focus on distributed scenarios and SFC availability. SPIDER is designed to make SFC placement in different dis- tributed scenarios, i.e., scenarios with different hardware and software characteristics. To do that, SPIDER uses context information in order to define the SFC placement strategy. In addition, machine learning techniques are used to predict the traffic of allocated SFCs and reinforcement learning to select the servers for the SFC placement. We compare the perfor- mance of LSTM and GRU models to predict traffic using a real dataset of cellular network. In order to define the placement of an SFC request, we proposed a reinforcement learning based algorithm to select the suitable candidate node and define the redundancy strategy to meet availability requirements. We implemented a proof-of-concept of SPIDER in order to show the feasibility of the framework. We implemented the SFCs using containers and Kubernetes to manage them. We assess the framework by assessing the placement time for SFCs with different numbers of VNFs. In order to evaluate the SFCs placed, we also evaluate the SFC delay for a centralized and a distributed scenario.