In–Kernel Fast Path Performance For Containers Running Telecom Workload


Nishanth Shyamkumar
Piotr Raczynski
Dave Cremins
Michal Kubiak
Ashok Sunder Rajan


Nuts and Bolts

Session Type




Virtualization of Telecommunication workloads opens the door to flexible and resource efficient deployment in a cloud infrastructure, for Telecom operators. The layers of virtualization involved in containers while providing advantages, also present a bottleneck for network functions requiring high performance. Relying on third party data-path libraries as a solution result in increased external dependencies, that affect ease of integration at the orchestration layer, and loss of Linux native debugging and diagnostic support.

In this paper, we explore and make the case to use an in-kernel based network pipeline for a 5G User Plane Function, that reduces the CPU core count from 10 to 1 for processing 2 Million Packets Per Second. This is achieved by implementing AF_XDP zero copy support in an SR-IOV Virtual Function driver. We demonstrate performance comparisons with alternate in-kernel data path mechanisms, as part of the evaluation. The proposed solution maintains performance at par with third party poll mode drivers, but importantly maintains operability with Linux tools, scalability and allows Kubernetes orchestration of the User Plane Function running in pods.