Precise Dynamic Resource Allocation for Virtualised Network Functions

Abstract The chaining and optimal placement of Virtual Network Functions (VNFs) in a static network has been extensively discussed and investigated. However a key trend of current network evolution is towards network softwarisation which is targeted at addressing the exponential growth in traffic demand that is causing progressive congestions in today’s Communication Service Providers (CSPs) networks. In network softwarisation, VNFs are being deployed as part of complex network service chains, containing multiple VNFs distributed across multiple points of presence (PoPs). VNFs are instantiated and resources have to be precisely allocated dynamically on-demand otherwise this can lead easily to either bad quality of service or waste of resources if too few or too many are allocated. To mitigate the challenges, we propose a Precise Dynamic Resource Allocation (PDRA) model which is a machine learning-based dynamic resource allocation targeted at adequately allocating VNF resources. The model has the following components: VNF profiling, VNF performance measurement, machine learning training and VNF auto scaling and placement. In comparison to standard fixed resource allocation, our model significantly addresses VNF under and over allocation of resources resulting in low latency and improved service quality.