Comparison of UPF acceleration technologies and their tail-latency for URLLC

The industry is striving to deploy private cellular networks in their factories. The envisioned use cases, such as industrial automation, require Ultra-Reliable and Low-Latency Communications (URLLC). In 5G, the User Plane Function (UPF) handles user traffic and hence significantly influences the performance in terms of latency and throughput. However, present UPFs are often user-space applications running on commercial off-the-shelf (COTS) servers introducing computational delays that follow long-tailed distributions. In this work, we compare modern packet processing acceleration technologies such as XDP, DPDK, and P4 and their respective tail-latency performance. Our results demonstrate the clear advantage of hardware-based P4 implementations in terms of reduced tail-latency of about 6μ8 at a percentile of 99.999% for small packets. Furthermore, we also demonstrate that software-based frameworks can compete with hardware-based approaches. For example, DPDK achieves tail-latencies of approximately 23μ8 at a percentile of 99.999 %. Based on our findings, we illustrate the trade-offs between low-cost software- and high-performant hardware-based packet processing systems.