BUFFERBLOAT, NOT THE RADIO: LATENCY IN A MULTI-USER 5G STANDALONE NETWORK
TL;DR
It is concluded that mitigating queue buildup, rather than improving the radio link, is the primary lever for latency control in this regime, and the implications for active queue management and learning-based RAN control are discussed.
Abstract
End-to-end latency, not raw throughput, increasingly limits the quality of experience in modern mobile networks. This paper presents an experimental characterization of the latency bottleneck in a fully software-defined 5G Standalone (SA) network built from srsRAN, Open5GS, and a ZeroMQ-based radio emulator. Across twelve deterministic scenarios (S1-S12) that independently vary the number of active users, the offered load, the radio quality, and the resource-management policy, we measure downlink throughput, round-trip time (RTT), and packet loss with second-level aggregation over repeated 60-second runs. Three findings stand out. First, scaling the cell from one user to five or seven triggers severe bufferbloat: mean RTT rises from roughly 0.59 s to 5-7 s, with a 95th-percentile reaching 13.4 s and an anomalous run peaking near 29 s, while packet loss remains at zero. Second, degrading the radio channel from good to poor conditions does not reduce throughput or raise latency, because MAC-layer adaptation and HARQ/ARQ fully absorb the radio impairment under deterministic load; the dominant latency factor is buffer saturation, not the air interface. Third, a static resource-management policy worsens latency by about 12% relative to no control at moderate load. We conclude that mitigating queue buildup, rather than improving the radio link, is the primary lever for latency control in this regime, and we discuss the implications for active queue management and learning-based RAN control.