AI 中文总结
该研究对比云原生5G核心网中UPF的四种I/O模式，基于SD-Core BESS-UPF实测性能，给出不同场景下的模式选择建议，还记录了部署陷阱。
AI 中文摘要
用户面功能（User Plane Function，UPF）承载5G网络中的所有用户面流量，其吞吐量取决于数据包在网卡（NIC）与应用之间的传输方式，即数据包I/O模式。我们在单个开源UPF上对比了四种广泛使用的模式：AF_PACKET、AF_XDP、云原生数据平面（Cloud Native Data Plane，CNDP）以及数据平面开发套件（Data Plane Development Kit，DPDK）。我们将SD-Core BESS-UPF以Charmed算子形式部署在Canonical Kubernetes中的Intel XXV710网卡上，在每种模式下运行相同的GTP-U/PDR/FAR/QER流水线，仅更改BESS端口驱动，因此所有差异均可归因于I/O后端。对于每种模式，我们描述了其架构、数据路径、内存模型和部署要求，并测量了吞吐量、延迟、CPU使用率和稳定性；还针对具有商用用户设备（UE）的分离式O-RAN 5G无线接入网（RAN）对部署进行了端到端验证。在基于RFC 2544标准的XXV710/i40e测试平台上，64字节数据包场景中，AF_PACKET达到0.25 Mpps，2个工作线程时CNDP为5.52 Mpps、AF_XDP为6.47 Mpps，4个工作线程时DPDK为10.30 Mpps，8个工作线程时DPDK为13.09 Mpps，且平均延迟最低（8.1微秒）。在匹配工作线程数量时，三种内核旁路模式的性能差异在测量误差范围内，且AF_XDP在每核心性能上领先；DPDK的优势源于扩展上限效应，因为其基于vfio-pci的原生PMD突破了绑定CNDP和AF_XDP的每网卡设备AF_XDP套接字限制（后两者被限制在2个工作线程），而非每数据包效率优势。当巨页（hugepages）、vfio-pci和隔离核心成本过高时，CNDP和AF_XDP是云原生场景的最优选择；DPDK则适用于专用主机。我们还记录了合成基准测试中未出现的部署陷阱，包括Kubernetes的某配置错误悄无声息地将DPDK吞吐量减半。本文为在云原生5G部署中选择UPF数据平面模式的运营商提供了并排参考依据。
英文摘要
The User Plane Function (UPF) carries all user-plane traffic in a 5G network, and its throughput depends on how packets move between the NIC and the application, that is, on the packet I/O mode. We compare four widely used modes, AF_PACKET, AF_XDP, the Cloud Native Data Plane (CNDP), and the Data Plane Development Kit (DPDK), on a single open-source UPF. Using SD-Core BESS-UPF deployed as a Charmed operator on Intel XXV710 NICs in Canonical Kubernetes, we run the same GTP-U/PDR/FAR/QER pipeline under each mode and change only the BESS port driver, so any difference is attributable to the I/O backend. For each mode we describe its architecture, datapath, memory model, and deployment requirements, and we measure throughput, latency, CPU usage, and stability; the deployment is also validated end-to-end against a disaggregated O-RAN 5G RAN with a commercial UE. On an XXV710/i40e testbed (NDR per RFC 2544) at 64 B, AF_PACKET reaches 0.25 Mpps, CNDP 5.52 and AF_XDP 6.47 Mpps at 2 workers, and DPDK 10.30 Mpps at 4 workers and 13.09 at 8, with the lowest latency (8.1 microseconds average). At matched worker counts the three kernel-bypass modes are within noise and AF_XDP leads per core; DPDK's advantage is a scaling-ceiling effect, since its native PMD over vfio-pci escapes the per-netdev AF_XDP socket limit that pins CNDP and AF_XDP at two workers, not a per-packet efficiency win. CNDP and AF_XDP are the cloud-native sweet spot when hugepages, vfio-pci, and isolated cores are unaffordable; DPDK is justified on dedicated hosts. We also document deployment pitfalls absent from synthetic benchmarks, including a Kubernetes limits.cpu mis-setting that silently halved DPDK throughput. The paper is a side-by-side reference for operators choosing a UPF dataplane mode in a cloud-native 5G deployment.
Comments34 pages, 19 figures
