knowledge centre

Data plane acceleration technologies: realising the potential of network virtualization

Gorkem Yigit Senior Analyst, Research
Caroline Chappell Research Director

"Acceleration technologies will be pivotal to network transformation by enabling operators meet the performance, latency, QoS, subscriber density and security requirements of existing and future applications with optimum TCO."


Data plane acceleration technologies: realising the potential of network virtualization

Network operators have virtualized many control plane network functions but they face technical limitations and cost disincentives when they migrate data plane functions to general-purpose network function virtualization infrastructures (NFVI) that are based on x86 platforms. Many acceleration technologies are available to address these challenges, from software-centric acceleration approaches to hardware-centric solutions. These can be used to offload compute-intensive virtual network function (VNF) workloads and reclaim virtualization overhead. For some use cases, such as AI/ML and video processing, graphics processing units (GPUs) and vision processing units (VPUs) can provide significant performance enhancements. However, operators are confused about the benefits of, and the dependencies between, each approach, as well as when to use each to support specific network functions and use cases.

This white paper aims to improve operators’ understanding and awareness of acceleration technologies in order to maximise the performance and efficiency of their NFV/SDN, 5G and AI/ML applications. The paper outlines the key drivers for acceleration technologies and identifies the main use cases and functions that require acceleration. It introduces a range of software-centric and hardware-centric acceleration technologies, analyses their advantages and challenges, as well as their suitability for different VNF and NFVI use cases. It also presents performance results for accelerating virtualized security functions. Finally, it discusses the need for an acceleration technology abstraction layer and a uniform management framework to simplify and automate the operations of heterogenous acceleration resources with an ‘as-a-service’ model. It highlights industry initiatives such as OpenStack Cyborg and OPNFV DPACC, which aim to help operators achieve these goals.