5G UPF Measurement Dataset: Comparative Forwarding Performance of OAI-UPF, eUPF, Free5GC, and Open5GS, and QoS Enforcement of the eBPF/XDP-Based OAI-UPF

This dataset accompanies the paper "The Brewing Storm in 5G's Data Plane: Design and Evaluation of a High-Performance eBPF/XDP-Based User Plane Function", submitted to IEEE Transactions on Network and ServiceManagement (ISSN 1932-4537), 2026. It contains the raw and processed measurements, the analysis scripts, and the complete test configurations used to evaluate the eBPF/XDP-based OAI User Plane Function (UPF) against three other open-source UPF implementations. ## Contents **1. Forwarding-performance campaign**Throughput, CPU utilisation, and packet loss rate (PLR) for four implementations: OAI-UPF, eUPF, Free5GC, and Open5GS. - Packet sizes: 64, 250, 500, 750, 1000, 1250, 1460 Bytes (downlink); 104, 250, 500, 750, 1000, 1250, 1500 Bytes (uplink)- Rx queues / CPU cores: 2 to 32 in steps of two- Injection rates: per-implementation, bounded by measured saturation- 1408 distinct configurations, ~235 hours of generated traffic **2. QoS-enforcement campaign** (OAI-UPF only)OAI-UPF is the only implementation among the four providing QER-baseddownlink shaping, so these measurements are single-implementation.- Intra-UE scenario: one UE, six concurrent QoS flows (one Non-GBR ICMP and five GBR flows with distinct GBR/MBR budgets)- Inter-UE scenario: 50 UEs joining and leaving in staggered batches, with ten reference UEs carrying the measurements **3. Control experiment**Traces from a bare XDP_REDIRECT path between N6 and N3, bypassing the UPFpipeline entirely, used to attribute small-packet losses to the NIC driverrather than to the implementations under test. **4. Virtualised comparison**Separate measurements against UPG-VPP, which could not be operated withthe NICs of the main testbed. ## Measurement setup Two identical hosts, Ubuntu 22.04 with Linux kernel 6.2, AMD EPYC 9534(64 cores, 2.45 GHz base / 3.55 GHz all-core boost, 256 MB L3). Trafficgenerated with Cisco TRex over dual-port Mellanox ConnectX-5 Ex NICs(100 Gbps per port); UPF host fitted with dual Intel E810 NICs(100 Gbps per port). Each forwarding configuration was measured over a single continuous 600 srun with per-second sampling; the first 60 s are discarded as warm-up.QoS scenarios run for approximately 1000 s with per-second iperf3 andping sampling. ## Known limitation The Intel E810 NICs exhibit an XDP TX buffer leak under sustained load,originating in the "ice" driver rather than in any UPF implementation.This primarily affects small-packet workloads (64 and 104 Bytes), whereper-queue packet rates are highest. Measurements at these sizes should beinterpreted accordingly. The control experiment in this datasetreproduces the same drop pattern with the UPF pipeline bypassed,confirming the driver-level origin. ## File organisation To be Updated ... ## Software The OAI UPF implementation is released separately under the OAI CommunitySoftware Source License:https://gitlab.eurecom.fr/oai/cn5g/oai-cn5g-upf(GitHub mirror: https://github.com/openairinterface/oai-cn5g-upf) ## License Data and documentation: CC BY 4.0. Analysis scripts: ⟨license⟩.