Part of: OCUDU on COSMOS · 4G / 5G & O-RAN tutorials
This page drives a commercial Foxconn RPQN-7801 O-RU from the permissively-licensed OCUDU DU (built on
srsRAN Project) over a real O-RAN 7.2 (split-7.2x) fronthaul — eCPRI on VLAN 4, S-plane PTP (G.8275.1) — with a native Open5GS core and a commercial Quectel 5G modem as the UE.
Status: working. A Quectel modem camps, registers and carries data over the Foxconn O-RU: TCP DL 176–181 / UL 136–144 Mbit/s, UDP DL 316 / UL 167 Mbit/s, ping 24 ms, DL BLER 2–6 % at MCS 20–21. Last verified 2026-10-26 on image
ocudu-20261020(OCUDU 26.10). Five settings differ from srsRAN's reference for this RU; they are listed under Results.
This is the production O-RAN shape on the OCUDU (srsRAN) platform: instead of the DU driving its
own USRP, the radio lives in a standalone Foxconn RPQN-7801 O-RU connected by an Open Fronthaul
7.2x link. OCUDU's Open Fronthaul (ru_ofh) block speaks eCPRI (enhanced Common Public Radio Interface) directly to the RU using a raw-socket
backend (no DPDK) over a VLAN-4 tagged interface, with the S-plane synchronised by G.8275.1 PTP (Precision Time Protocol)
against the fabric grandmaster.
The Foxconn RU runs in file mode (a local RRHconfig_xran.xml + an activation script), not M-Plane
NETCONF, so the RU is configured out-of-band and the DU streams C/U-plane to it. Its S-plane is restarted by hand after every RU reset, its C/U-plane is activated from a shell on the RU, and a hardware watchdog resets it if the sequence stalls, so the bring-up is one tight sequence. The UE side has two Foxconn-specific points: the real subscriber key (it is not the generic test key) and the modem band lock (n78 at 3.75 GHz enumerates as band 77). PLMN is the testbed standard 00101, and the UE pool is 192.168.100.0/22.
For the radio-less software version see the OCUDU soft-only
tutorial; for the USRP over-the-air version
see the OCUDU cross-vendor
tutorial. The same OCUDU DU over
the Benetel RAN650 and LiteOn FlexFi is covered by the OCUDU O-RAN 7.2
tutorial, and this O-RU driven by OCUDU with NVIDIA's GPU layer 1 on a GH200 by OCUDU on NVIDIA Aerial.
PTP lock, C/U-plane activation, UE camp (RSRP −74, SINR 23), registration, PDU session and data all work.
| CPU platform | Image | TCP DL | TCP UL | UDP DL | UDP UL | Ping | Date |
|---|---|---|---|---|---|---|---|
| B — 2× Intel Xeon Gold 6126 (24 cores / 48 threads) | ocudu-20261020 (OCUDU 26.10) |
176–181 | 136–144 | 316 | 167 | 24 ms avg (min 12.6) | 2026-10-26 |
| B — 2× Intel Xeon Gold 6126 (24 cores / 48 threads) | ocudu-20261020 (OCUDU 26.10), fresh load, this page followed step by step |
157–167 | 143–162 | 327–330 | 167–168 | 22–24 ms avg (min 11.5) | 2026-10-26, 3 runs |
Mbit/s, iperf3 receiver line, 10 s, one stream, Quectel RM520N-GL. UDP DL is capped by the modem's USB 2.0 link (24 % loss at 450 Mbit/s offered), not by the cell.
Settings that differ from srsRAN's reference for this RU, and why
enable_dl_static_compr_hdr: true — the RU's file sets RRH_CMPR_HDR_PRESENT = 0, so the DL U-plane must carry no compression header (eCPRI payload = 12 header bytes + 7644 IQ bytes for 273 PRBs). With a header the RU reads the IQ offset, trips AD9025 PA protection error and no UE sees the cell.enable_ul_static_compr_hdr: true — the RU sends no UL compression header either; with false the DU drops every UL symbol.expert_phy.max_proc_delay — at 10 the DU discards about 6000 DL U-plane messages per second as late (tx_kpis nof_late_up_dl); at the default, 0.xRN c_late); at 285–470, 0.Run exactly one ptp4l on the RU: its ptp4l.service restarts itself 30 s after it is killed, and two instances fighting over the clock make the RU drop out 1–5 min after the DU starts. Stop the unit, then run one run_ptp4l.sh (Step 4).
Facts that hold for this RU
root/root). Port 22 is always refused; that does not mean the RU application failed.ptp4l → lock → cuplane → one DU start with no gap.c_on climbing in /var/log/rrh_log_print/rrh_log_print.log (grep xRN); core1 is running N is not.foxconn-oru1 (3650–3850 MHz, so the 3750 MHz carrier is bookable). With that reservation active, power it from your domain console with omf tell -a {on,offh,reset} -t foxconn-oru1.instrument.orbit-lab.org.The DU config and the bring-up scripts are in the COSMOS project repository, configs/ocudu-foxconn-72/: gnb-ocudu-foxconn.yml, fx-host-prep.sh (DU host), fx-ru-latch.sh (RU), fx-ocudu-run.sh (RU latch + one DU start), fxa.sh (the RU shell helper both of them call; copy it into the same directory), and on the UE node fx_at.py (one AT command), fx-ue-data.sh (PDU session) and fx-ue-perf.sh (the throughput runs).
This page exercises two fronthaul ideas explained on O-RAN and the 7.2x fronthaul: eAxC identifiers, which label each fronthaul stream (here DL/UL on 0–3 and PRACH on 4–7) and must agree between the DU config and the RU's RRHconfig_xran.xml; and O-RU management modes — the Foxconn is configured from a file on the RU rather than by the DU over NETCONF, which is why the DU config has to mirror the RU file by hand.
Figure — The C-plane and U-plane share VLAN 4 and the S-plane comes from the fabric grandmaster; each DL, UL and PRACH stream is an eAxC lane whose number must match on both ends.
After completing this tutorial you will be able to:
rrh_log_print c_on counter (not the misleading "core1 is running" line).Authentication failure (MAC failure) as a wrong-key symptom.| Learning path | Optional, after ocudu-72 — O-RAN 7.2 with a commercial O-RU. Previous: OCUDU with a commercial O-RU · Next: OCUDU with Aether SD-Core |
| Status | Working: UE attach + data (2026-10-26) |
| Difficulty | Advanced |
| Estimated time | 75–120 min |
| Domain / sandbox | an osc E810 fronthaul server + the Foxconn O-RU (instrument domain) |
| Topic group | Cellular (4G/5G/O-RAN) |
| Last verified | 2026-10-26 end to end (attach, ping, TCP/UDP both directions) on ocudu-20261020 |
Background knowledge
ssh/ip/iperf3 on Linux.Account & access
Devices / nodes
| Resource | Role | Qty | Notes |
|---|---|---|---|
E810 fronthaul server (osc or sb1) |
OCUDU OFH DU + Open5GS core | 1 | fronthaul NIC on VLAN 4, iface DATA1a (e.g. srv18.osc) |
| Foxconn RPQN-7801 O-RU | commercial O-RU (radio) | 1 | foxconn-oru1.instrument.orbit-lab.org, RU shell/M-Plane SSH 10.3.16.1:830 (port 22 is always refused), RU MAC 6c:ad:ad:00:0d:60 |
| Quectel RM520N-GL | commercial 5G UE | 1 | USB on a UE node; AT /dev/ttyUSB2/3, QMI wwan0/cdc-wdm0 |
Disk images
| Image | Load onto | Provides |
|---|---|---|
ocudu.ndz |
DU node | OCUDU gnb (/opt/ocudu), Open5GS (/opt/open5gs), UHD/qmicli, cosmos-ofh-isolation (the image carries no isolcpus — they are derived per host, see Setup) |
lte5gue.ndz |
UE node | qmicli (libqmi), modem helper scripts |
(ocudu.ndz / lte5gue.ndz are symlinks to the latest dated builds.)
Software components
| Component | Version | Source |
|---|---|---|
| OCUDU gNB (srsRAN Project) | 26.10 e0db566 (image ocudu-20261020) |
/opt/ocudu/build/apps/gnb/gnb — no license needed, built with ru_ofh raw-socket |
| Open5GS 5GC | build-tree | /opt/open5gs, mongo db open5gs, launcher /root/open5gs/start.sh |
linuxptp (ptp4l/phc2sys) |
distro | S-plane G.8275.1 on the DU host + the RU |
| Foxconn RU firmware | v3.1.15q.551v0706-oam |
file-mode activation init_rrh_config_enable_cuplane |
Spectrum / RF / special
dl_arfcn 650000 (3750 MHz),udCompHdr in either direction), eAxC IDs 0, 2 (DL), 0–3 (UL) + 4–7 (PRACH),00101, slice SST 1 / SD 000001, DNN internet (IPv4), UE pool192.168.100.0/22, gateway 192.168.100.1, internet-NAT'd.Shared RF space. The
sb1fronthaul rig shares one physical RF space, so a Quectel modem on any
co-located node can camp on this cell. If a UE "sees no signal", the RU is not radiating — it is never
an antenna/EIRP problem.
Figure — The DU server's fronthaul NIC reaches the O-RU through the fabric switch, which is also the PTP grandmaster; power control of the O-RU is a separate path through omf tell on your fronthaul console.
The OCUDU DU + Open5GS core run on one E810 compute node; its fronthaul NIC carries the VLAN-4 eCPRI
link to the Foxconn O-RU, which radiates over the air to the Quectel modem:
DU / core node (osc) Foxconn RPQN-7801 O-RU Quectel UE
┌────────────────────────┐ eCPRI ┌───────────────────────┐ n78 OTA ┌──────────────┐
│ OCUDU gnb (ru_ofh) │ VLAN 4 │ file-mode RRHconfig │ 3.75 GHz │ RM520N-GL │
│ Open5GS 5GC │═════════════►│ eAxC 0-3 DL/UL, 4-7 RA │────────────►│ wwan0 QMI │
│ ogstun 192.168.100.1 │ + PTP S-pl │ M-Plane 10.3.16.1:830 │ RSRP -68 │ 192.168.100.2│
│ AMF 127.0.0.5:38412 │◄────────────►│ TX 24 dBm, 4T4R │ │ │
└────────────────────────┘ G.8275.1 └───────────────────────┘ └──────────────┘
DATA1a (VLAN 4) PTP GM = fabric switch (G.8275.1)
127.0.0.5:38412 (the N2 interface).ogstun (192.168.100.1) and is NATed to the Internet.DATA1a carries C/U-plane on VLAN 4; eAxC 0–3 are the DL/UL antenna streams and 4–7 the PRACH streams.10.3.16.1:830; nothing else uses that path.OCUDU needs no license server and no DPDK — the ru_ofh block uses a raw AF_PACKET socket on the
named interface.
ssh <username>@console.<domain>.cosmos-lab.org
-r 0):omf load -i ocudu.ndz -t <du-node> -r 0 -o 2400 -b 600 # a large image: without -r 0 the load reports failure at 90 %
omf load -i lte5gue.ndz -t <ue-node> -r 0 -o 2400 -b 600
omf tell -a on -t <du-node>,<ue-node>
sshd on a host without CPU isolation. The ocudu.ndz image deliberately carries no isolcpus; derive and apply them for the fronthaul NIC DATA1a with cosmos-ofh-isolation --apply, reboot, and run the checks exactly as in Real-time host preparation — Command 1. Expected on srv18.osc (2026-09-14): isolated CPUs 3-11,27-35 (nine physical cores and their hyperthread siblings), NTP service: inactive, and one phc2sys -s DATA1a -w -n 24 -m, settled to a small offset (s2, tens of ns). Use -w, not -O 0 — see Troubleshooting.irqbalance, lift RT throttling, put the fronthaul NIC in promiscuous mode, run ptp4l (G.8275.1) on DATA1a plus phc2sys, and start the Open5GS core. fx-host-prep.sh from configs/ocudu-foxconn-72/ in the COSMOS project repository does all of it, with pidfiles in /run/fx (fx-host-prep.sh stop undoes it); it includes the core launcher of Step 1 below, so with it Step 1 is already done:ssh root@<du-node>
bash /root/fx-host-prep.sh
FX-HOST-READY
This tutorial is a manual bring-up: the RU activation is driven from a shell on the RU, with a human watching the PTP lock. Do the steps in order: core →
DU config → RU activate → gNB → verify radiating → attach UE. The core and subscriber come first because they do not depend on the RU; the RU comes last before the DU because its watchdog punishes any gap between activation and the DU start.
The launcher starts all Open5GS network functions, creates the ogstun gateway interface and installs NAT to the Internet.
ssh root@<du-node>
/root/open5gs/start.sh # NFs + ogstun 192.168.100.1/22 + NAT→internet
ip -4 addr show ogstun | grep inet
inet 192.168.100.1/22 scope global ogstun
The core authenticates the SIM with MILENAGE, which needs the same secret key (Ki) and operator key (OPc) on both sides. If the core holds a different key, the UE rejects the network at the first authentication.
ocudu.ndz already carries the COSMOS SIMs, including the osc Quectel's (read its IMSI with AT+CIMI, or from cosmos-sim-db), so check first and provision only a card that is missing. The image holds two rows for some IMSIs; that is harmless. The COSMOS test SIMs are not programmed with the generic srsRAN test key. Each card's real
Ki/OPc is in the image's SIM database, cosmos-sim-db, which provisions the COSMOS SIMs into the core;
read your card's values from there (for your own SIM, use its values).
(The key values are left out of this page on purpose.)
Provision it into Open5GS with opc set and no op_type (a stray op_type breaks MILENAGE), plus
the slice and DNN:
mongosh open5gs --quiet --eval '
db.subscribers.updateOne({imsi:"001010000000032"},
{$set:{"security.k":"<Ki>",
"security.opc":"<OPc>",
"security.op":null,"security.amf":"9001",
"slice.0.sst":1,"slice.0.sd":"000001","slice.0.session.0.name":"internet"},
$unset:{"security.op_type":""}})'
Wrong-key symptom. If the AMF logs
Authentication failure(MAC failure)on a fresh attempt
(preceded byUnknown UE by SUCI), the network key ≠ the SIM's — re-read the values fromcosmos-sim-db(above).
A cached-GUTI re-registration can succeed with a wrong key; aCFUN=0/1on the modem forces the
real MILENAGE test.
Because the Foxconn is configured from its own file rather than by the DU, the DU config has to mirror that file: carrier frequency, bandwidth, MAC addresses, VLAN and eAxC numbers must match value for value. Each comment below names the RU-side key it must equal.
The gNB config carries the cell + a ru_ofh block matched to the RU's RRHconfig_xran.xml. Use configs/ocudu-foxconn-72/gnb-ocudu-foxconn.yml in the COSMOS project repository
(copy it to /root/ocudu/ on the DU node — its expert_execution CPU sets are srv18's, re-derive them
for another host). The fields that must agree with the RU:
cell_cfg:
dl_arfcn: 650000 # n78, 3750 MHz (== RU RRH_LO_FREQUENCY_KHZ 3750000)
band: 78
channel_bandwidth_MHz: 100 # RU RRH_MAX_PRB 273
common_scs: 30
plmn: "00101"
tac: 7
pci: 3 # unique per O-RU (Benetel 2, LiteOn 1)
nof_antennas_dl: 2 # RU branches 1 and 3 (2 and 4 read ~6 dB low on this unit)
nof_antennas_ul: 4
ru_ofh:
t1a_max_cp_dl: 470 # srsRAN reference 250-420 makes ~28 % of C-plane late at the RU (xRN c_late)
t1a_min_cp_dl: 285
t1a_max_cp_ul: 470
t1a_min_cp_ul: 285
t1a_max_up: 196
t1a_min_up: 80
ta4_max: 1960 # srsRAN reference 500; same value as for the LiteOn
ta4_min: 25
is_prach_cp_enabled: true
ignore_ecpri_payload_size: true
compr_method_ul: bfp
compr_bitwidth_ul: 9
compr_method_dl: bfp
compr_bitwidth_dl: 9
enable_ul_static_compr_hdr: true # == RU RRH_CMPR_HDR_PRESENT 0: false drops every UL symbol
enable_dl_static_compr_hdr: true # == RU RRH_CMPR_HDR_PRESENT 0: no udCompHdr on the DL; false = PA protection trips, no cell
iq_scaling: 1.0
cells:
- network_interface: DATA1a # raw socket, no DPDK
ru_mac_addr: 6c:ad:ad:00:0d:60 # == RU RRH_SRC_MAC_ADDR
du_mac_addr: 00:11:22:33:44:66 # == RU RRH_DST_MAC_ADDR
vlan_tag_cp: 4
vlan_tag_up: 4
prach_port_id: [4, 5, 6, 7] # == RU RRH_EAXC_ID_TYPE3
dl_port_id: [0, 2] # RU branches 1 and 3 (RRH_EAXC_ID_TYPE1 lists 0-3)
ul_port_id: [0, 1, 2, 3]
# no expert_phy.max_proc_delay: at 10 the DU drops ~6000 DL U-plane messages/s as late
This step brings the RU from a fresh reset to "PTP locked and C/U-plane enabled". It has to happen with the gNB stopped, and Step 5 has to follow at once.
The RU's shell is its M-Plane SSH on port 830 (legacy KEX, root/root; port 22 is always refused).
Start every attempt from a fresh RU reset or power cycle, from your domain console while your foxconn-oru1 reservation is active:
omf tell -a reset -t foxconn-oru1.instrument.orbit-lab.org # or offh, then on, for a cold cycle
Then, on the RU (from the DU node via the M-Plane SSH), replace the RU's ptp4l with exactly one instance and run the cuplane activation:
ptp4l.service rather than only killing ptp4l: the unit restarts it 30 s later, and two ptp4l fighting over the clock make the RU drop out minutes after the DU starts.state=0, rms ~100 ns). A second restart locks in seconds.Wait until /tmp/ptp4l_user_gen.cfg exists (a few minutes after boot). The gNB must be down for this, and go straight on to Step 5 — the RU watchdog-resets itself if it is left idle:
ssh -p 830 -o KexAlgorithms=+diffie-hellman-group14-sha1 \
-o HostKeyAlgorithms=+ssh-rsa root@10.3.16.1
# on the RU:
ls /tmp/ptp4l_user_gen.cfg # must exist first
systemctl stop ptp4l.service; killall ptp4l run_ptp4l.sh 2>/dev/null
(setsid /usr/utils/run_ptp4l.sh > /tmp/ptp.log 2>&1 &)
pidof ptp4l # exactly one PID
grep -a state= /tmp/ptp.log | tail -1 # wait for state=3, rms <= 30 ns; if it hunts, repeat the two lines above once
ptp4l[...]: state=3 rms 6 max 8 freq +60 delay 9 # SLAVE, rms a few ns
The killall above runs inside the O-RU's own embedded Linux, which serves one DU at a time; it stops only the RU's PTP daemons. Do not use killall on COSMOS servers. /tmp/ptp.log matters: the cuplane activation below waits on that file for the lock.
What to look for: state=3 is the linuxptp SLAVE state — the RU's clock is following the grandmaster — with rms in single-digit nanoseconds.
# still on the RU — stop the activation started at boot (before PTP locked), then enable C/U-plane (reads RRHconfig_xran.xml):
killall init_rrh_config_enable_cuplane 2>/dev/null
cd /home/root/test && (./init_rrh_config_enable_cuplane > /tmp/L.out 2>&1 &)
grep -aE 'valid|core1' /tmp/L.out | tail -3
Setting of LO1 and PRB is valid
Setting of Tx attenuation is valid
core1 is running 9
What to look for: the two valid lines — the RU accepted the carrier (LO1/PRB) and transmit attenuation from its file. core1 is running is not proof of radiation; that is checked in Verification.
The ~15-minute RU resets. The RU runs a hardware watchdog (
rrh_hw_watchdog) and resets itself
roughly every 10–15 minutes when it is left idle or mid-setup (PTP not yet locked, no DU streaming).
Only a tight reset → PTP → cuplane → DU sequence avoids them. The optional loop below keeps
rrh_log_print.logsmall; it does not affect the resets:setsid sh -c 'while true; do :> /var/log/rrh_log_print/rrh_log_print.log; sleep 15; done' &
configs/ocudu-foxconn-72/fx-ocudu-run.sh <gnb-cfg>(in the COSMOS project repository) runs the RU part of
this step (fx-ru-latch.sh) and Step 5 as one sequence on the DU node, right after the RU reset; it refuses to
start a second gNB.fx-cycle.shstops the gNB again (SIGINT through its pidfile).
The DU is started once, detached from the SSH session so it survives a dropped connection, and under a CPU mask that includes the isolated cores.
With the RU PTP-locked and cuplane activated, start the gNB once on the isolated cores (keep stdin
open, detach from ssh). Do not restart it while the RU stays latched — to retry, go back to the RU reset
in Step 4:
ssh root@<du-node>
cd /opt/ocudu/build
nohup setsid bash -c 'sleep 1800 | taskset -c 0-47 \
./apps/gnb/gnb -c /root/ocudu/gnb-ocudu-foxconn.yml > /tmp/g.log 2>&1' & disown
sleep 30; pgrep -x gnb >/dev/null && echo ALIVE; grep -a 'to AMF completed' /tmp/g.log
ALIVE
N2: Connection to AMF on 127.0.0.5:38412 completed
taskset -c 0-47 (all cores on a 48-core host — use 0-$(( $(nproc --all) - 1 )) elsewhere) is required:
isolcpus hides the isolated cores from srsRAN's affinity validator, so the process must be launched with
a mask that includes them (see the main-pool section of Real-time host preparation).
What to look for: ALIVE (the process survived its first 30 s) and N2: Connection to AMF ... completed — the gNB has registered with the core. Neither says anything yet about the radio.
| Check | Command / where | Expected |
|---|---|---|
| RU PTP | grep -a state= /tmp/ptp.log on the RU |
state=3, rms a few ns, pidof ptp4l = exactly one PID |
| DU ↔ core | grep -a 'to AMF completed' /tmp/g.log |
N2: Connection to AMF on 127.0.0.5:38412 completed |
| RU radiating | grep -a xRN /var/log/rrh_log_print/rrh_log_print.log on the RU |
c_on climbing, c_late=0; err_tci/ecpri/port/sct all 0; no new PA protection lines (the log keeps lines from before the reset, dated before the RU's clock synced) |
| UE camped | AT+QENG="servingcell" |
NR5G-SA, 001,01, ARFCN 647328 |
| Registered | AMF log | Registration complete for the IMSI |
| Data | ping -I $WWAN |
0 % loss to 192.168.100.1 and 8.8.8.8 |
All six rows passed on 2026-10-26.
c_on, not "core1 is running"The authoritative "is it transmitting" signal is the RU's rrh_log_print counter: c_on climbing with
err_tci/ecpri/port/sct all zero (the srsRAN Foxconn success signature):
# on the RU:
grep -a xRN /var/log/rrh_log_print/rrh_log_print.log | tail -2
xRN: total=734319 c_early=0 c_on=66948 c_late=0 err_tci=0 err_ecpri=0 err_port=0 err_sct=0 err_total=648520
xRN: total=750079 c_early=0 c_on=68372 c_late=0 err_tci=0 err_ecpri=0 err_port=0 err_sct=0 err_total=662457
What to look for: c_on (C-plane messages that arrived on time) growing between the two lines, with c_early and c_late at 0 — the DU's messages are landing inside the RU's receive window.
On the UE node, hard-lock the Quectel to NR5G-SA and force the PLMN. Set band 77:78, not 78 —
3.75 GHz enumerates as band 77, so a bare 78 lock blocks the cell. The general modem procedure is on COTS UE: the Quectel modem; the band lock is the Foxconn-specific part.
ssh root@<ue-node> 'systemctl stop ModemManager'
# via AT on the Quectel's if02 port, /dev/serial/by-id/usb-Quectel_*-if02-port0 (python3 fx_at.py '<cmd>' sends one):
AT+CFUN=0
AT+QNWPREFCFG="mode_pref",NR5G
AT+QNWPREFCFG="nr5g_disable_mode",0
AT+QNWPREFCFG="nr5g_band",77:78
AT+CRSM=214,28539,0,0,12,"FFFFFFFFFFFFFFFFFFFFFFFF" # clear forbidden-PLMN list
AT+CFUN=1
AT+COPS=1,2,"00101",12
AT+QENG="servingcell"
+QENG: "servingcell","NOCONN","NR5G-SA","TDD",001,01,00066C000,1,7,647328,78,3,-68,-10,30,1,71
NR5G-SA, PLMN 001/01, RSRP -68, SINR 30 (2026-08-19; on 2026-10-26 RSRP −74, SINR 23). Confirm registration in the AMF log:
ssh root@<du-node> "grep -a 'Registration complete' /tmp/open5gs/amf.log | tail -1"
[gmm] INFO: [imsi-001010000000032] Registration complete
This is the rig's variant (APN 'internet' through qmi-proxy with -p, a /30 address) of the data bring-up explained under Bringing up data; add only scoped routes through wwan0, as shown there.
ssh root@<ue-node>
# the Quectel's own QMI device + netdev; never assume cdc-wdm0/wwan0: a UE node can carry two modems
AT=$(ls /dev/serial/by-id/usb-Quectel_*-if02-port0 | head -1) # /dev/quectel-at where the image has it
D=$(readlink -f /sys/class/tty/$(basename $(readlink -f $AT))/device/../..)
QMI=/dev/$(ls $D/*/usbmisc); WWAN=$(ls $D/*/net); echo "Quectel: $QMI $WWAN"
ip link set $WWAN down; echo Y > /sys/class/net/$WWAN/qmi/raw_ip; ip link set $WWAN up # raw_ip only changes while $WWAN is DOWN
qmicli -d $QMI -p --wds-start-network="apn='internet',ip-type=4" --client-no-release-cid
qmicli -d $QMI -p --wds-get-current-settings | grep -E 'IPv4 address|gateway'
IPv4 address: 192.168.100.2
IPv4 gateway address: 192.168.100.1
QMI CID-alloc timeout? Reload the driver:
rmmod qmi_wwan; modprobe qmi_wwan(with ModemManager
stopped), then retry--wds-start-network.
AT=$(ls /dev/serial/by-id/usb-Quectel_*-if02-port0 | head -1) # /dev/quectel-at where the image has it
D=$(readlink -f /sys/class/tty/$(basename $(readlink -f $AT))/device/../..)
QMI=/dev/$(ls $D/*/usbmisc); WWAN=$(ls $D/*/net)
ip addr add 192.168.100.2/30 dev $WWAN; ip link set $WWAN up mtu 1400
ping -c5 -I $WWAN 192.168.100.1 # UPF gateway
ping -c5 -I $WWAN 8.8.8.8 # internet via UPF NAT
5 packets transmitted, 5 received, 0% packet loss # to the gateway
5 packets transmitted, 5 received, 0% packet loss # to 8.8.8.8 via the UPF NAT
Ping to the gateway averaged 24 ms (min 12.6 ms, 0 % loss) on 2026-10-26.
Throughput — run an iperf3 server on the UPF gateway (192.168.100.1, DU node) and drive it from the UE bound to its PDU address, 10 s per direction (-R for downlink), as in Bringing up data on COTS UE: the Quectel modem. Receiver lines on 2026-10-26 (ocudu-20261020):
| Test | Receiver line |
|---|---|
| TCP DL | 176–181 Mbit/s (two runs) |
| TCP UL | 136–144 Mbit/s (two runs) |
| UDP DL | 316 Mbit/s at 450 Mbit/s offered — the modem's USB 2.0 link is the limit |
| UDP UL | 167 Mbit/s |
DL BLER stayed at 2–6 % at MCS 20–21. A COTS UE reaching the internet over a commercial O-RU driven by a license-free DU.
Stop processes as process groups identified by their executable — never with pkill -x or pkill -f, which match by name or command line and can hit another user's process or your own shell. Define this helper in your shell on each node you log into:
stop_exe() { # signal the process GROUP of every process executing exactly $1 (default TERM)
local exe sig=${2:-TERM} d st pgid
exe=$(readlink -f "$1")
for d in /proc/[0-9]*; do
[ "$(readlink "$d/exe" 2>/dev/null)" = "$exe" ] || continue
st=$(cat "$d/stat" 2>/dev/null) || continue
st=${st##*) } # drop "pid (comm) ", safe for any comm
pgid=$(echo "$st" | awk '{print $3}') # remaining fields: state ppid pgrp
kill -"$sig" -- "-$pgid" 2>/dev/null
done
}
The function identifies a process by the binary it is executing (/proc/<pid>/exe), which a command line cannot fake, and signals its whole process group. The gNB was launched with setsid, so its group is exactly that launch. It sends TERM first; if something is still running after a few seconds, repeat the call with KILL as the second argument. /usr/bin/iperf3 matches every iperf3 on the node — on your reserved nodes that is only yours.
ssh root@<ue-node>
stop_exe /usr/bin/iperf3
AT=$(ls /dev/serial/by-id/usb-Quectel_*-if02-port0 | head -1) # /dev/quectel-at where the image has it
D=$(readlink -f /sys/class/tty/$(basename $(readlink -f $AT))/device/../..)
QMI=/dev/$(ls $D/*/usbmisc); WWAN=$(ls $D/*/net)
qmicli -d $QMI -p --wds-stop-network=disable; ip addr flush dev $WWAN
ssh root@<du-node>
stop_exe /opt/ocudu/build/apps/gnb/gnb
stop_exe /usr/bin/iperf3
# from your domain console: power off the nodes and the O-RU
omf tell -a offh -t <du-node>,<ue-node>
omf tell -a offh -t foxconn-oru1.instrument.orbit-lab.org
Do the O-RU power-off while your foxconn-oru1 reservation is still active.
Two failures apply to any O-RU driven by srsRAN's Open Fronthaul library and are silent — every RAN counter reads healthy while nothing works: a fronthaul NIC that is not promiscuous (the DU's AF_PACKET socket then receives zero frames while tcpdump still shows them; see Real-time host preparation — promiscuous mode), and gains too high for bench range (the UE is metres from the O-RU, and vendor defaults assume a real cell radius). Both are in the table below.
Warning: Do not run a DPD (digital pre-distortion) reset on a radio that is already working: on the RAN650 it leaves transmit branches without output until a full power cycle.
| Symptom | Likely cause | Fix |
|---|---|---|
Modem stuck SEARCH, no SSB seen |
RU not radiating | Check c_on climbing (Verification); if not, stop the gNB and re-do Step 4 from the RU reset, then start the gNB once — a DU restart under a latched RU does not revive it |
RU unpingable at 10.3.16.1 |
the RU's hardware watchdog reset it (idle / mid-setup), or it wedged | omf tell -a reset -t foxconn-oru1.instrument.orbit-lab.org (or offh then on) from your domain console; then re-latch without gaps |
ssh root@10.3.16.1 → Connection refused |
port 22 is never open on this RU | use the M-Plane SSH on port 830 (legacy KEX, root/root) |
RU PTP never reaches state=3, or hunts at state=0 with rms ~100 ns |
a restart before /tmp/ptp4l_user_gen.cfg existed (failed to open configuration file); or the first restart after a cold boot only stepped the RU's stale clock |
on the RU (its embedded Linux, not a COSMOS server): wait for the cfg file, stop ptp4l.service, start one run_ptp4l.sh, and repeat that once if it hunts (Step 4) |
RU drops out (tx_recover, then unpingable) 1–5 min after the DU starts; the RU journal shows announce timeouts and selected local clock as best master |
two ptp4l on the RU: its ptp4l.service restarted itself 30 s after a hand killall |
systemctl stop ptp4l.service before starting one run_ptp4l.sh; check pidof ptp4l shows one PID |
DU drops every UL symbol: dropped received Open Fronthaul message as no section was decoded correctly |
UL compression-header mismatch — this RU firmware sends no UL compression header | enable_ul_static_compr_hdr: true (srsRAN's RPQN reference has false) |
RU log repeats AD9025 PA protection error about once a second, tx_recover, no UE sees the cell; lowering iq_scaling changes nothing |
the DU sends a DL compression header the RU does not expect | enable_dl_static_compr_hdr: true |
| DL BLER 40–80 % at every MCS with 4 DL ports; the RU reports TX branches 2 and 4 ~6 dB below 1 and 3 | weak TX branches 2 and 4 on this unit | dl_port_id: [0, 2], nof_antennas_dl: 2 — BLER 2–6 %, TCP DL ~180 |
RU xRN c_late a large share of C-plane messages (~28 %) |
C-plane arrives too late for the RU at srsRAN's reference T1a 250–420 µs | T1a C-plane windows 285–470 µs → c_late = 0 |
DU tx_kpis nof_late_up_dl in the thousands per second |
expert_phy.max_proc_delay: 10 set |
remove max_proc_delay |
gNB: Invalid CPU core selected '8' |
isolcpus hides cores |
launch inside taskset -c 0-<ncpu-1>; see Real-time host preparation |
OFH woke up late … skipped N symbols |
the host clock and the O-RU disagree | NTP off, then phc2sys -s DATA1a -w -n 24 (offset s2, tens of ns). Use -w, not -O 0: the G.8275.1 PHC is on the TAI timescale, so -O 0 leaves CLOCK_REALTIME 37 s ahead of a UTC-keeping O-RU and the two derive System Frame Numbers 3700 frames apart. See Real-time host preparation |
ether_rx rx_bytes = 0, no PRACH, tcpdump shows the frames |
the fronthaul NIC is not promiscuous | ip link set <fh-iface> promisc on — see Real-time host preparation |
DL BLER stuck 20–30 % while dl_olla is already several dB negative |
iq_scaling too high: the DU's IQ clips before compression, which link adaptation cannot back away from |
lower iq_scaling (5.5 at bench range on the Benetel) |
Uplink a few kbit/s, one RX port at exactly 0.000000 dBFS |
O-RU receive branch saturated by its own uplink gain | ul_gain_correction=0 in the RU config, then reboot the RU (read only at boot) |
rx_late a large fraction of rx_total, RTT hundreds of ms |
uplink reception window too tight | ta4_max: 1960 |
| ptp4l never reaches SLAVE and the log fills with management messages | a PTP management-message storm on the fabric | drop them at ingress (tc filter … protocol 0x88f7 … match the management msgType … action drop) and raise announceReceiptTimeout — the full egress filter is on Real-time host preparation |
AMF Authentication failure(MAC failure) on a fresh attempt |
wrong network Ki/OPc | use the cosmos-sim-db key (Step 2); do not set op_type |
Modem camps commercial LTE (LIMSRV,LTE) |
mode_pref drifted to AUTO |
re-assert mode_pref=NR5G and COPS=1,2,"00101",12; see COTS UE: the Quectel modem |
qmicli … CID allocation failed |
ModemManager holds the QMI device / stuck channel | systemctl stop ModemManager; rmmod qmi_wwan; modprobe qmi_wwan; see COTS UE: the Quectel modem |
ping: Network is unreachable |
wwan0 has no IP yet |
wait for --wds-get-current-settings to return an address, then ip addr add |
On this wiki
External
ptp4l, phc2sys): https://linuxptp.sourceforge.net/Author(s): COSMOS team · Last verified: 2026-10-26 end to end (attach, ping, TCP/UDP both directions) · Tested image/release: ocudu-20261020 (OCUDU 26.10 e0db566); Foxconn RPQN-7801 (v3.1.15q.551v0706-oam) · Tags: ocudu, srsran, oran, 7.2, foxconn, o-ru, quectel, open5gs, promiscuous, gain