Computer Science editorial
Comparing UPF Dataplane I/O Modes in a Cloud-Native 5G Core: AF_PACKET, AF_XDP, CNDP, and DPDK on SD-Core BESS-UPF
The core problem
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. Operators deploying cloud-native 5G cores must choose among several packet I/O backends, each with distinct architecture, memory model, and deployment requirements. This paper compares 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.
The authors use SD-Core BESS-UPF deployed as a Charmed operator on Intel XXV710 NICs in Canonical Kubernetes. They 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 they describe its architecture, datapath, memory model, and deployment requirements, and they 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.
The central research question is practical: which packet I/O mode should an operator choose for a cloud-native 5G UPF, and under what conditi
Innovation
On the XXV710/i40e testbed at 64-byte packets under NDR per RFC 2544, the four modes show sharply different throughput profiles. AF_PACKET reaches 0.25 Mpps. CNDP reaches 5.52 Mpps at 2 workers, and AF_XDP reaches 6.47 Mpps at 2 workers. DPDK reaches 10.30 Mpps at 4 workers and 13.09 Mpps at 8 workers, with the lowest latency at 8.1 microseconds average.
At matched worker counts, the three kernel-bypass modes—AF_XDP, CNDP, and DPDK—are within noise of one another, and AF_XDP leads per core. This is a key finding: DPDK's headline advantage is not a per-packet efficiency win but a scaling-ceiling effect. Its native PMD over vfio-pci escapes the per-netdev AF_XDP socket limit that pins CNDP and AF_XDP at two workers. In other words, DPDK wins by allowing more workers to run effectively, not by processing each packet faster.
The latency result reinforces this interpretation: DPDK's 8.1 microseconds average is the lowest measured, but the gap at matched worker counts is small. The stability measurements and CPU usage data further support the conclusion that the kernel-bypass modes are broadly comparable in per-core efficiency, with deployment constraints—hugepages, vfio-pci, isolated
Why it matters
The results reframe the common assumption that DPDK is simply faster than kernel-bypass alternatives. At matched worker counts, AF_XDP, CNDP, and DPDK are within noise, and AF_XDP leads per core. DPDK's advantage is a scaling-ceiling effect: its native PMD over vfio-pci escapes the per-netdev AF_XDP socket limit that pins CNDP and AF_XDP at two workers. This distinction matters for operators because it changes the decision criteria from raw throughput to deployment feasibility.
The authors conclude that 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 where those resources can be provisioned. This is a practical taxonomy: kernel-bypass modes differ less in per-packet efficiency than in their operational prerequisites.
The paper also documents deployment pitfalls absent from synthetic benchmarks, including the Kubernetes limits.cpu mis-setting that silently halved DPDK throughput. Such findings are valuable because they reflect real operator experience rather than idealized lab conditions. The end-to-end validation against a disaggregated O-RAN 5G RAN with a commercial UE further strengthens the external validity of the comparison.
For operators, the decision framework can be summarized as follows:
The broader implication is that cloud-native 5G core performance is as much an orchestration and resource-isolation problem as a packet-processing problem. The paper serves as a side-by-side reference for operators choosing a UPF dataplane mode in a cloud-native 5G deployment.
Who should read this
Opening member content…