Skip to content

Exploring GPU Acceleration in free5GC: Registration and User-Plane Processing

Note

Author: Chia-Hui Chen、Yu-Chen Chan
Date: 2026/10/11


1. Introduction

Can GPUs improve performance in existing free5GC network functions? We explore this question through two experiments: accelerating database access during user equipment (UE) registration, and accelerating packet processing in the User Plane Function (UPF). Together, they examine how GPU execution and CPU/GPU cooperation affect registration completion time and user-plane processing rates.

2. Two GPU Acceleration Experiments

2.1. Exploring GPU-Resident Databases for the 5G Core

Explores whether a GPU-resident database can reduce UE registration time in free5GC. We implement a document service that replaces MongoDB access in UDR, NRF, PCF, CHF, and BSF while preserving existing service interfaces and network topology.

The CPU handles request parsing, query planning, and scheduling, while the GPU evaluates document predicates and accesses data resident in VRAM. The service uses two read banks for asynchronous execution and a global read/write barrier to maintain operation ordering. We compare this prototype against a separate free5GC deployment using native MongoDB, with common performance optimizations aligned.

Experiments launch one UE every millisecond across populations of 100 to 10,000 UEs, with ten measured runs per configuration. Timing extends from the first registration start to the last recorded Registration Accept, without PDU session establishment. Mean differences are modest at smaller populations, but the GPU configuration reduces the mean recorded registration window by approximately 25.0% at 5,000 UEs and 49.0% at 10,000 UEs.

These results suggest potential benefits for sustained registration workloads. Persistence and recovery are outside the prototype’s scope, and further profiling is needed to identify the sources of improvement.

2.2. GPU Packet Processing for the UPF Data Plane

We explore GPU acceleration of packet processing in free5GC's UPF through a CPU/GPU forwarding backend. The CPU handles network input/output, control and reporting, while the GPU parses packets, selects forwarding rules and prepares output packets. Functional checks completed 3,000 UE-to-data-network echo roundtrips through the prototype.

We also implemented GPU processing of usage counters and tunnel sequence numbers. To reduce repeated searches across a packet batch, the GPU builds ordered lists of packets contributing to each forwarding or usage rule. Each rule then processes its own list, preserving the order needed for state updates.

Using rules captured from ten real sessions and controlled packets matching them, we compared memory-resident processing on an RTX 3090 with an Intel i9-12900 using up to eight performance cores. Inputs were already in each processor's memory, and rate-based policing was disabled. CPU led the exploratory uplink and mixed-traffic cases at the largest tested CPU budget. Longer paired downlink trials measured approximately 16–21% higher GPU processing rates than the corresponding eight-core CPU configurations. These results identify a workload-dependent processing benefit from the optimized GPU implementation.

3. Conclusion

The two experiments examine different opportunities for GPU acceleration in free5GC. The document-service prototype reports shorter registration windows at larger UE populations, while the UPF experiment identifies GPU processing gains in selected downlink workloads. CPU remains competitive, including its lead in the UPF uplink and mixed-traffic comparisons.

The results measure different outcomes: registration completion time and memory-resident packet processing rates. The UPF processing gains do not establish a live forwarding speedup. Database persistence and recovery, together with UPF network I/O, batch waiting, latency and energy costs, are important considerations when extending these prototypes toward operational use.

Reference