Performance summary

Mirrored documentation

Mirrored read-only from claudia-5gc @ v2.6.0. Edit upstream, not here.

Headline user-plane throughput of the ClaudIA 5GC UPF. These are lab figures measured against the environment below; the method, raw results and tuning findings are part of the members documentation.

Single UPF instance, optimized build, Docker Desktop VM (10 vCPUs), 1400 B packets, ≥ 99 % delivered, 0 UPF drops:

Traffic	Max throughput	Packets/s	UPF CPU
Downlink	≈ 1.8 Gb/s (99 % delivered up to ≈ 2.2 Gb/s)	160–198 k	≈ 95–120 % of a core
Uplink	≥ 2.7 Gb/s (generator-limited floor)	≥ 244 k	≈ 100 % of a core
Both at once	≈ 3.6 Gb/s combined (1.8 up + 1.8 down); degrades at 4.4 Gb/s	160 k each way	≈ 1.5 cores
64 B, both at once	≥ 400 kpps combined (≈ 200 Mb/s; ladder-limited floor)	200 k each way	≈ 1.7 cores

One UERANSIM UE: ≈ 102–115 Mb/s (UE-bound, linear to 4 UEs ≈ 410 Mb/s). Not claimed: carrier-grade numbers, > 16 sessions, rate-limited (AMBR) traffic at scale (single-UE enforcement results are in the notes below), real-RAN behaviour. These are lab figures against the environment manifest below.

Environment: Docker Desktop on an Apple host (10 vCPUs, 8 GB). Figures are valid only for that setup. The “knee” is the highest offered load with ≥ 99 % delivery and zero UPF drops; “floor” means the traffic generator, not the UPF, was the limit.

Reproduce: make perf-up, then make perf-bench (UPF in isolation, synthetic sessions), make perf-dn / make perf-multi-ue (real UEs through the UPF). Knobs: n3.socket_buffer_bytes, n6.tun_mtu, n6.tcp_mss_clamp in the UPF config — see ../configuration.md.