Posting this because I can't find any published Flint 4 throughput data on the
10G/SFP+ path — every review I found tested on 1–2.5 Gbps links and explicitly
noted the 10G ports and SFP+ cage went untested. So this may be a gap worth
filling, and I'd like to know whether others on multi-gig see the same shape.
Short version: a LAN client behind the router tops out around 2.2–2.9 Gbps to
the internet, while the router itself pulls ~5 Gbps from the same server on the
same port. But running both at once totals 7.84 Gbps — so neither the line nor
the router's aggregate capacity is the limit. Hardware acceleration is on and
binding flows, and the CPU is idle throughout.
Setup
| Router | GL-BE14000 (Flint 4), firmware 4.9.1 |
| OS | OpenWrt 21.02-SNAPSHOT, kernel 5.4.281, mediatek/mt7988 |
| WAN | XGS-PON ONU stick in the SFP+ cage, VLAN 131 tagged, DHCP (10 Gbps plan) |
| LAN client | Fedora 44, Realtek RTL8127 10GbE (r8169), link 10000Mb/s full, MTU 1500 |
| Test target | speedtest.init7.net (Zurich, ~1 ms away), iperf3 -P 8 -t 12 -R |
Network Acceleration (HNAT) enabled. DPI disabled. No QoS/SQM. No VPN.
Results
Alternating runs, same server, same port, back to back:
| Round | Router itself | LAN client (forwarded) |
|---|---|---|
| 1 | 5.20 Gbps | 2.48 Gbps |
| 2 | 4.91 Gbps | 2.26 Gbps |
| 3 | 5.09 Gbps | 2.18 Gbps |
Reproducible ~2x gap. Then the interesting one — both at once, different
server ports:
router : 4.53 Gbits/sec
client : 3.31 Gbits/sec
-----------------------
total : 7.84 Gbits/sec
Two things to note: the WAN clearly sustains at least 7.84 Gbps, and the
forwarded client ran faster under contention (3.31) than it ever does alone
(~2.3). A simple forwarding bottleneck shouldn't behave that way.
Stream count helps a little but doesn't close the gap:
client solo, -P 8 : 2.24 Gbits/sec
client solo, -P 16 : 2.91 Gbits/sec
client solo, -P 32 : 2.67 Gbits/sec
What I ruled out
- Cabling —
ethtool -Serror counters all zero on both hosts, hundreds of GB
passed; both links negotiate 10000Mb/s full duplex. - Router CPU —
topduring a forwarded transfer shows 97% idle. - Hardware acceleration not engaging —
/sys/kernel/debug/hnat/hnat_entry
shows the flows bound during the transfer (PPE1 BIND cnt = 9,
PPE2 BIND cnt = 8for 8 iperf streams), NAT entries visible in both
directions through the VLAN 131 WAN.mtkhnat.global.enable='1'. - DPI interference — stopping
netifydchanged nothing (1.98 → 2.06 Gbps). - Client NIC / MTU — client was originally on MTU 9000, which forced
r8169
to disable TSO and TX checksum offload and had oversized frames dropped by the
router. Fixing it to 1500 raised the client's LAN transmit from 1.24 → 3.47
Gbps, but did not change the forwarded-internet figure materially. - Test server — the router pulls ~5 Gbps from the same server and port, and
the simultaneous run shows 7.84 Gbps aggregate, so the server isn't the cap.
Solo forwarded runs do show meaningful retransmits (3,000–4,400 per 12–15 s run),
which the router's own runs also show at higher throughput — so loss alone doesn't
explain the ordering.
Questions
- Is ~2.2–2.9 Gbps expected for a single forwarded client on the Flint 4's
10G path, or should it track the ~5 Gbps the router achieves itself? - Does anyone else running the SFP+ cage as a tagged-VLAN WAN (common for
PON/ISP setups) see the same? I'd like to rule out the VLAN tag interacting
badly with the offload path. - Is there a known reason the NPU would prioritise CPU-terminated traffic over
hardware-forwarded traffic? A published Flint 4 review documented a related
"cross-port forwarding vs CPU-termination asymmetry" on the same firmware,
though in that case it only appeared under simultaneous load — mine happens
with a single idle-router client.
Happy to run any additional tests or share raw iperf3/HNAT/top output. Reproducing
this needs nothing exotic: two hosts, iperf3 to a nearby public server, and
comparing the router's own result against a LAN client's.