Design and optimization of a DPU-driven software and programmable hardware collaborative architecture for UPF

As mobile networks evolve, the user plane function (UPF) faces demands for high bandwidth, low latency, and massive connectivity. However, skewed traffic distribution and diverse services lead to a surge in fine-grained rules. The resulting rule dependencies have become a critical bottleneck for UPF forwarding performance. Moreover, existing algorithms for resolving dependencies suffer from high computational latency and fail to meet strict timing requirements. To address this, a data processing unit (DPU) driven software-hardware hyper UPF architecture was proposed. This architecture separated elephant and mice flows based on traffic features. Elephant flows were offloaded to programmable hardware for acceleration. In contrast, the CPU handled mice flows and employed a fast independent rule generation algorithm to eliminate rule dependencies. Additionally, a rule storage structure and multi-pipe cooperative update mechanism were proposed. Experimental results show that under a high concurrency test environment, the system achieved a throughput of 97 Gbit/s with an end-to-end latency below 500 µs. The computational overhead for rule dependency resolution was reduced by over 65%, with rule storage and update efficiency improved by approximately 50%.