Li, Xicheng (2026) Accelerating network services with programmable data planes. PhD thesis, University of Glasgow.
Full text available as:|
PDF
Download (1MB) |
Abstract
Programmable data planes allow network functions to be offloaded onto the forwarding path, potentially avoiding the latency and coordination overhead of processing packets in control-plane or end-host software. The open challenge is to identify the resource and processing costs, and the restrictions on service behaviour, required for a network function to operate within the data plane without runtime control-plane or user-space intervention, and how to retain a fast packet-processing path when its logic requires cross-packet history or external context.
This thesis investigates this question through two contrasting arrangements of runtime state maintenance, and uses a framework based on state ownership and data-plane independence to compare their costs and constraints.
Programmable Dataplane (PDP) technologies like P4 and Extended Berkeley Packet Filter (eBPF) each have inherent limitations: P4's pipeline model, which varies by target, constrains arithmetic operations and state management, and may impose per-packet execution budgets; eBPF, while more flexible as it runs on the end host, is subject to verifier-imposed execution constraints.
To address these challenges, the thesis designs and evaluates a Domain Name System (DNS) cache implemented entirely in a P4 pipeline on the Behavioral Model v2 (BMv2) software switch, and develops an eBPF/eXpress Data Path (XDP) architecture for IEC 61850 Generic Object Oriented Substation Event (GOOSE) anomaly detection in Industrial Control Systems (ICS) with user-space coordination and digital twin integration.
The DNS cache performs lookup and update without per-packet control-plane intervention, serving hits locally and forwarding misses or refresh queries upstream. On BMv2, it supports bounded-length A-record lookups, uses recirculation, and allocates approximately 78 KiB of register state for cache and epoch storage. An indicative comparison reports approximately 32 % fewer parser states visited per packet with binary-decomposition padding removal on recirculation, compared with a byte-loop implementation. Its admission and freshness trade-offs illustrate the service restrictions of maintaining state entirely within the data plane.
The ICS architecture validates GOOSE sequence progression in XDP, delegates anomaly aggregation and deny-list updates to user space, and provides a co-simulation interface to physical-process context. In the evaluated Virtual Ethernet (veth) testbed, the XDP detection path has a median per-packet latency of 3.815 μs, compared with 2.766 μs for a no-op XDP baseline. Mean reaction intervals from the threshold-triggering anomaly to the first subsequently blocked frame are 116 and 127 ms for two sequence-violation attack classes, including event delivery, the user-space map update, and the wait for the next frame. These results support the feasibility of the detection and enforcement paths.
Taken together, the prototypes show that software-maintained policy can support an independent packet-processing path, while maintaining state entirely within the data plane imposes resource costs and service restrictions. These trade-offs help guide the selection of an offloading mechanism.
| Item Type: | Thesis (PhD) |
|---|---|
| Qualification Level: | Doctoral |
| Colleges/Schools: | College of Science and Engineering > School of Computing Science |
| Supervisor's Name: | Pezaros, Professor Dimitrios and Li, Dr. Emma |
| Date of Award: | 2026 |
| Depositing User: | Theses Team |
| Unique ID: | glathesis:2026-86369 |
| Copyright: | Copyright of this thesis is held by the author. |
| Date Deposited: | 18 Nov 2026 16:18 |
| Last Modified: | 18 Nov 2026 16:19 |
| Thesis DOI: | 10.5525/gla.thesis.86369 |
| URI: | https://theses.gla.ac.uk/id/eprint/86369 |
| Related URLs: |
Actions (login required)
![]() |
View Item |
Downloads
Downloads per month over past year

Tools
Tools