Dynamic service chain composition in virtualised environment

Network Function Virtualisation (NFV) has contributed to improving the flexibility of network service provisioning and reducing the time to market of new services. NFV leverages the virtualisation technology to decouple the software implementation of network appliances from the physical devices on which they run. However, with the emergence of this paradigm, providing data centre applications with an adequate network performance becomes challenging. For instance, virtualised environments cause network congestion, decrease the throughput and hurt the end user experience. Moreover, applications usually communicate through multiple sequences of virtual network functions (VNFs), aka service chains, for policy enforcement and performance and security enhancement, which increases the management complexity at to the network level. To address this problematic situation, existing studies have proposed high-level approaches of VNFs chaining and placement that improve service chain performance. They consider the VNFs as homogenous entities regardless of their specific characteristics. They have overlooked their distinct behaviour toward the traffic load and how their underpinning implementation can intervene in defining resource usage. Our research aims at filling this gap by finding out particular patterns on production and widely used VNFs. And proposing a categorisation that helps in reducing network latency at the chains. Based on experimental evaluation, we have classified firewalls, NAT, IDS/IPS, Flow monitors into I/O- and CPU-bound functions. The former category is mainly sensitive to the throughput, in packets per second, while the performance of the latter is primarily affected by the network bandwidth, in bits per second. By doing so, we correlate the VNF category with the traversing traffic characteristics and this will dictate how the service chains would be composed. We propose a heuristic called Natif, for a VNF-Aware VNF insTantIation and traFfic distribution scheme, to reconcile the discrepancy in VNF requirements based on the category they belong to and to eventually reduce network latency. We have deployed Natif in an OpenStack-based environment and have compared it to a network-aware VNF composition approach. Our results show a decrease in latency by around 188% on average without sacrificing the throughput.