Please note: This master’s research paper presentation will take place in DC 2310.
Gan
Wang,
Master’s
candidate
David
R.
Cheriton
School
of
Computer
Science
Supervisor: Professor Martin Karsten
Network Function Virtualization (NFV) facilitates the implementation of network functions in software, enabling them to run in virtual machines on standardized hardware from different vendors. Recently, the serverless paradigm has emerged as a promising scheme for supporting virtual network functions due to the paradigm’s inherent scalability, security, and reduced operational overheads. Serverless platforms automatically allocate resources for incoming requests, which reduces the need for manual intervention during runtime. Additionally, leveraging serverless platforms provides access to robust security features and best practices established by cloud providers. Economically, they promote a pay-as-you-use paradigm, ensuring costs are incurred only for actual compute time, which makes them particularly suitable for applications with fluctuating workloads. These characteristics suggest that serverless paradigms can potentially support data plane network functions. However, previous studies indicate that serverless computing is not a naturally practical solution for data plane network functions. This is because of its lack of persistency. In serverless platforms, every function invocation requires retaining states from an external database, which generates high latency.
This paper studies how to adjust serverless platforms so that they can provide more efficient support for data plane network functions while preserving traditional valuable serverless features, especially security features. This paper focuses on serverless packet forwarding. It proposes leveraging in-kernel packet processing and maintaining ephemeral local states on each node to improve forwarding efficiency. Also, this paper introduces a P4-eBPF compilation module to add another layer of security to the system. A prototype focusing on the 5G User Plane Function (UPF) use case is presented to validate the feasibility of deploying data plane functions based on the proposed node design.