01 · Brief
The question: how does enabling 5G Quality-of-Service change observable network behaviour for different traffic types, and can the difference be measured on a fully open-source testbed? The answer was a complete digital prototype built on a single Ubuntu 24.04 host — srsRAN Project gNodeB, srsRAN 4G UE over a ZeroMQ virtual radio, Open5GS 5G Core and the O-RAN Software Community Near-RT RIC — with no radio hardware at all.
02 · Build
On top of the stack we wrote a Python xApp that subscribes to the gNB's E2SM-KPM service model and writes per-second, per-bearer telemetry labelled by 5QI and slice, a multi-profile UDP traffic generator covering seven traffic classes, and six orchestration scripts so the whole experiment can be brought up and repeated with one command. Two matched campaigns — best-effort and QoS-aware scheduling — produced a balanced dataset of 4,111 one-second samples, analysed in pandas.
03 · Status
Three findings: QoS costs 2–4% of mean throughput in single-UE conditions; throughput variability roughly doubles under QoS, contradicting the intuition that guaranteed-bit-rate shaping smooths traffic; and zero-throughput samples rise four to six times, direct evidence of admission control at work. The testbed is a reproducible base for closed-loop machine-learning xApps, and the same skills — Linux, containers, telemetry pipelines, statistical analysis — are what we bring to cloud and data work for clients.