Part of: Duranta on COSMOS · 4G / 5G & O-RAN · Step 3 of the Duranta learning path · Previous: duranta-ota
Split the Duranta base station at the O-RAN 7.2 fronthaul. The same nr-softmodem binary and OAI CN5G core run on a COSMOS server, now as an O-DU, and drive a commercial Benetel RAN650 O-RU over an Ethernet fronthaul instead of a USRP. The DU configures the O-RU itself over the O-RAN management plane (M-Plane) and keeps that session open while the cell runs. A commercial Quectel modem then registers, gets an IP address and carries real traffic over a 100 MHz n78 cell.
This is the production shape of a 5G base station, and the shape operators deploy. The 100 MHz, 4-antenna cell is far wider than the USRP cell of step 2, and the fronthaul is a standard interface. It also brings a new kind of problem: the O-RU is a closed, shared box that keeps whatever the last user wrote into it, and a wrong setting does not show up as an error. Most of the work in this tutorial is putting the O-RU into a known state before the DU starts, and proving that it is in it.
Concepts
- O-DU / O-RU — 3GPP splits a gNB into a central unit (CU) and a distributed unit (DU); the O-RAN Alliance adds the radio unit (RU), which holds the lowest physical layer and the RF. Here
nr-softmodemis CU and DU in one process (the O-DU), and the Benetel is the O-RU.- 7.2x fronthaul — the O-RAN split inside the physical layer: frequency-domain IQ carried as eCPRI over Ethernet. Duranta's transport is
fhi_72, built on the xran library, running with DPDK on an SR-IOV virtual function (VF) of an Intel E810 NIC.- Planes — the C/U-plane (control and IQ data), the S-plane (PTP time: both ends lock to the fronthaul switch, G.8275.1) and the M-Plane (NETCONF over SSH, used to configure the O-RU).
- Golden O-RU state — the full O-RU configuration captured when this rig last worked. Every bring-up restores it and refuses to start the DU on a mismatch.
- Cold cycle — a real power-off of the O-RU from the console, as opposed to a reboot inside the O-RU.
| Link | Between | Carried over |
|---|---|---|
| C/U-plane | O-DU ↔ O-RU | eCPRI over an E810 VF, through the fronthaul switch |
| S-plane | switch → O-DU and O-RU | PTP (G.8275.1); the fronthaul switch is the grandmaster |
| M-Plane | O-DU ↔ O-RU | NETCONF over SSH; nr-softmodem is the client and holds the session for the life of the cell |
| Air | O-RU ↔ modem | band n78, 100 MHz (273 PRB), 30 kHz subcarrier spacing, TDD DDDSU, on the channel of your O-RU reservation |
| N2 / N3 / N6 | O-DU ↔ core ↔ internet | as in duranta-ota: SCTP and GTP-U over the Docker bridge demo-oai, NAT in oai-ext-dn |
| Power | console → O-RU | omf tell under your O-RU reservation |
Where the O-RU's address and channel come from. You never type the O-RU's address. The tutorial names the O-RU by its COSMOS name (duranta_oru: benetel), and with duranta_ru_ip: auto and duranta_ru_name: auto it reads the O-RU's management address and its omf tell name from cosmos-oru-tools on the O-DU server when the run starts (the Benetel can answer on one of two management addresses; the tool picks the one that answers). The cell's band, centre and bandwidth come from your O-RU reservation (duranta_center_freq_mhz: reservation), so the cell is on the spectrum you booked. "auto" and "reservation" therefore mean "look it up at run time". To pin an address or a channel, set duranta_ru_ip, duranta_ru_name, duranta_center_freq_mhz or duranta_bw_mhz in your --params file; a channel you set must stay inside your reservation.
| Domains | rrail, osc, sb1 |
| Nodes | 2, named in OMF_NODES |
| On the console | python3, ansible-playbook, ssh, scp |
cosmos-tutorial check duranta-72 verifies, on each node:
Not in the package - you provide:
You will also need:
No licence is involved.
One O-RU on the air at a time. The Benetel and the LiteOn share one RF space: only one of them may be powered, and never two DUs on one O-RU. The bring-up refuses to start when the other O-RU is on a co-channel carrier or another host holds an M-Plane session with the Benetel.
On the console of the domain your O-RU reservation targets, name the O-DU server first and the UE node second (a UE node in another domain takes its full name), and ask whether the tutorial can run. The check is read-only.
export OMF_NODES="<du-node>,<ue-node>"
cosmos-tutorial check duranta-72
It ends with RUNNABLE or NOT RUNNABLE and the reasons. It verifies that the O-DU server carries the Duranta build, an Intel E810, the O-RU management key and cosmos-oru-tools listing the Benetel, that the UE node has a modem, and that no other DU or gNB is running. If an image is missing, load it first:
omf load -i duranta.ndz -t <du-node> -r 0 -o 2400 -b 600
omf tell -a on -t <du-node>
export OMF_NODES=<node>,<node>
cosmos-tutorial check duranta-72
cosmos-tutorial run duranta-72 # ansible
cosmos-tutorial run duranta-72 --mode manual # manual
cosmos-tutorial jupyter duranta-72 # jupyter
cosmos-tutorial run duranta-72 --mode parity # parity
cosmos-tutorial run duranta-72 --mode teardown # teardown
The tutorial ships the same experiment three ways. All three read one parameter file, use one log parser, and write the same JSON result into ~/cosmos-runs/duranta-72/<UTC>-<mode>/ on the console. OMF_NODES is required and ordered: the O-DU server first, the UE node second. The O-RU is not in OMF_NODES; it is an instrument, powered from this console under your reservation. A first run takes 90–150 minutes, most of it the O-RU's five-minute cold boots and the one-time isolation reboot.
The playbook resolves the O-RU from cosmos-oru-tools, ships the portable host layer to the O-DU server, stages the O-RU for Duranta (M-Plane on, the per-stack profile and the golden's stored configuration, all of which load at the next cold boot) and cold-cycles it. It applies CPU isolation if the running kernel lacks it (one reboot per image load), runs the bring-up, and if the bring-up exits 12 it cold-cycles the O-RU and runs it again, up to twice. It then requires the cell in service and the DU process alive, reads the fronthaul counters, and runs the modem, data-call and measurement steps of the manual mode.
The manual mode runs the commands below for you. Run them yourself to see what each part does.
Step 1 — Find the O-RU (on the console). Its management address and omf tell name come from cosmos-oru-tools on the O-DU server.
D=$(cosmos-tutorial which duranta-72)
RU_IP=$(python3 $D/cosmos_addr.py oru benetel --du <du-node>)
RU_NAME=$(python3 $D/cosmos_addr.py oru benetel --du <du-node> --field omf); echo "$RU_NAME $RU_IP"
Step 2 — Isolate CPU cores (on the O-DU server, once per image load). A 7.2 DU busy-polls its timing threads; on a host without isolation it starves everything else, including sshd, within seconds. The tool derives the cores from the fronthaul NIC's NUMA node.
ssh root@<du-node>
cosmos-ofh-isolation --nic DATA1a --apply && systemctl reboot
# after the reboot:
grep -o 'isolcpus=[^ ]*' /proc/cmdline
Step 3 — Cold-cycle the O-RU. A commercial O-RU initialises its radio and fronthaul timing correctly only from a genuine power-off. omf tell needs your active O-RU reservation; confirm by ping that the O-RU really went down, because the reply alone does not prove it.
# on the console:
omf tell -a offh -t $RU_NAME
# on the O-DU server: wait until it stops answering
for i in $(seq 1 30); do ping -c1 -W1 $RU_IP >/dev/null 2>&1 || break; sleep 2; done
# on the console, 30 s later:
omf tell -a on -t $RU_NAME
# on the O-DU server: wait for its management address (about 70-80 s into the boot)
for i in $(seq 1 84); do ping -c1 -W2 $RU_IP >/dev/null 2>&1 && break; sleep 5; done
Step 4 — Ship the host layer (on the console). The package carries the parts that make the bring-up portable: oai72-host.py derives the VF, the NUMA node, the real-time cores and the PTP cores from the server itself; the bring-up script, the Benetel's per-stack profile, the channel helper and the golden tool go with it.
ssh root@<du-node> "mkdir -p /root/oru-profiles /root/oru-setup"
scp $D/host/oai72-host.py $D/host/oru_rf.py root@<du-node>:/root/
scp $D/host/benetel-bringup.sh root@<du-node>:/root/benetel-bringup.sh
scp $D/host/oru-profiles/oru-apply.py $D/host/oru-profiles/benetel-ran650.yml root@<du-node>:/root/oru-profiles/
scp $D/host/oru-setup/oru_golden.py $D/host/oru-setup/oru_creds.py root@<du-node>:/root/oru-setup/
Step 5 — Bring up the O-DU (on the O-DU server). One script does the whole host and O-RU sequence, each stage printed as === [n] ===: stop any previous DU, put the O-RU in M-Plane mode, assert the per-stack O-RU settings and its TDD pattern, restore the golden's stored configuration ([0e]), set the performance governor, create the VFs and bind one to DPDK, move the NIC interrupts off the isolated cores, reserve hugepages, start PTP, start OAI CN5G, write the AMF's address and this host's cores into the DU configuration, select the M-Plane fronthaul library, launch nr-softmodem, and finally check the whole O-RU state against the golden again after the DU's own M-Plane commit ([7]).
PF=DATA1a ORU_IP=$RU_IP ORU_OMF=$RU_NAME /root/benetel-bringup.sh
echo "exit $?"
Exit 12 means the O-RU held another configuration (another user's, or one written while it was off) and the script has just written Duranta's; the O-RU loads it only at a cold boot. Repeat Step 3 and run the bring-up again; two rounds can be needed (the profile, then the golden). Then confirm the cell and start the iperf3 server in the data-network container (copy duranta-72_dn_iperf.sh from $D):
grep -a 'is in service' /tmp/du_benetel.log
grep -a 'o-du 0\]\[rx' /tmp/du_benetel.log | tail -1 # fronthaul receive counters
DN=$(docker inspect -f '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' oai-ext-dn)
bash /root/duranta-72_dn_iperf.sh 5201
Step 6 — Attach the modem and measure (on the UE node). This is the modem sequence of duranta-ota with this cell's band mask (77:78, which must include the cell's band). Copy duranta-72_modem.py and duranta-72_iperf.sh from $D first; the iperf wrapper retries a test whose control connection drops.
ssh root@<ue-node>
systemctl stop ModemManager
read QMI WWAN <<<"$(python3 /root/duranta-72_modem.py resolve auto auto)"
python3 /root/duranta-72_modem.py configure /dev/ttyUSB2 00101 77:78 oai
python3 /root/duranta-72_modem.py wait-reg /dev/ttyUSB2 240 00101
read QMI WWAN <<<"$(python3 /root/duranta-72_modem.py resolve auto auto)" # configure may have reset the modem
python3 /root/duranta-72_modem.py call /dev/ttyUSB2 $QMI $WWAN oai 240 00101 | tee /tmp/wds.txt
read QMI WWAN <<<"$(python3 /root/duranta-72_modem.py resolve auto auto)" # the call recovery can reboot it
UE=$(awk -F': ' '/IPv4 address/{print $2}' /tmp/wds.txt | tr -d ' '); GW=$(awk -F': ' '/IPv4 gateway address/{print $2}' /tmp/wds.txt | tr -d ' ')
ip addr flush dev $WWAN; ip addr add $UE/22 dev $WWAN; ip link set $WWAN mtu 1400 up
ip route replace 192.168.100.0/22 dev $WWAN
for t in $DN 8.8.8.8; do ip route replace $t/32 via $GW dev $WWAN; done
ping -I $WWAN -c 10 -i 0.3 -W 2 $DN
ping -I $WWAN -c 10 -i 0.3 -W 2 8.8.8.8
/root/duranta-72_iperf.sh $DN 5201 $UE 10 -R # DL
/root/duranta-72_iperf.sh $DN 5201 $UE 10 # UL
The notebook runs on the console and drives both nodes over SSH. It performs the same steps and plots downlink and uplink throughput. It can also be run non-interactively with papermill, which writes an executed copy to a new file.
The parity mode runs Ansible, manual and Jupyter one after the other with the same nodes, O-RU and modem, and compares their results. It passes when every mode has the cell on air and the modem registered on the same cell, the fronthaul error counters are exactly 0, local ping loss is at most 20 %, downlink is in 20–1500 and uplink in 1–600 Mbit/s, and the modes agree within 60 % (downlink) and 80 % (uplink). A 7.2 fronthaul that drops anything is not healthy, which is why that counter must be zero rather than small. The harness runs the teardown on every exit.
The bring-up prints each stage. A healthy run passes both golden gates and ends with the DU alive and the cell in service:
GOLDEN-PERSISTENT-OK
...
DU-ALIVE
--- cell/AMF:
[NR_RRC] cell PLMN 001.01 Cell ID ... is in service
GOLDEN-OK
===== READY =====
The DU prints is in service before the fronthaul initialises, so a DU that then dies at DPDK start-up still shows it; DU-ALIVE and steadily increasing receive counters are what prove the cell is real. The modem then registers (+C5GREG: 0,1 on 001/01) and gets an address in 192.168.100.0/22. A healthy Benetel cell gives roughly 200–225 Mbit/s downlink and 65–75 Mbit/s uplink with a ping of about 31 ms. A parity run ends with PARITY: PASS:
=== duranta-72 parity ===
mode ue_ipv4 rsrp_dbm sinr_db fronthaul_errors ping_rtt_ms ping_remote_loss_pct dl_mbps ul_mbps
ansible 192.168.100.2 -66 21 0 31.36 0.0 194.0 60.2
manual 192.168.100.2 -65 21 0 32.06 0.0 222.0 65.0
jupyter 192.168.100.2 -65 21 0 31.16 0.0 206.0 53.0
dl_mbps spread max/min = 1.14 (tol 1.60)
ul_mbps spread max/min = 1.23 (tol 1.80)
PARITY: PASS (same experiment in all modes; all checks green)
(Real output from 2026-10-23, 1× AMD EPYC 9355P.)
Change any of these with cosmos-tutorial run duranta-72 --params my.yml. An unknown key or a wrong type is refused. cosmos-tutorial params duranta-72 prints the file.
| Key | Default | Meaning |
|---|---|---|
duranta_oru |
benetel |
Which O-RU drives the cell: a cosmos-oru name, benetel | liteon. |
duranta_ru_name |
auto |
The O-RU's omf tell name. auto = from cosmos-oru-tools (orus.json) on the O-DU host; a value pins it. |
duranta_ru_console |
`` | "" = run omf tell on THIS console. |
duranta_ru_ip |
auto |
The O-RU's management address. auto = the address that answers from the O-DU host, from cosmos-oru-tools (orus.json; the Benetel moves between two); a value pins it. |
duranta_ru_key |
auto |
the Benetel's SSH key on the O-DU host. auto = BENETEL_KEY in the O-RU logins the run pushes from |
duranta_ru_cycle |
True |
A COLD cycle is mandatory before every run: :830 must take ~300 s to reopen. ~80 s means the RU never powered down -- xran then comes up "RU0 mu0" (15 kHz under a 30 kHz PHY) and the RU never radiates. ⚠️ Learned 2026-09-08: the RU's NETCONF :830 is CLOSED while it is in standalone (non-M-Plane) mode, so a readiness check that polls :830 before the bring-up waits forever. |
duranta_ru_off_wait_s |
30 |
|
duranta_ru_netconf_timeout_s |
420 |
|
duranta_fh_iface |
DATA1a |
Per-run files on the node live in the run's own directory $COSMOS_NODE_RUN_DIR (/var/tmp/cosmos/, set by cosmos-tutorial; /var/tmp/cosmos/duranta-72.current points at the latest run). |
duranta_log_options |
level,nocolor,utc_time,thread_id |
OAI log options passed to nr-softmodem (--log_config.global_log_options). "utc_time" = a UTC time-of-day stamp ([hh:mm:ss.uuuuuu], CLOCK_REALTIME) on every line (owner decision 4: the EE capture aligns the DU log with the other roles). |
duranta_core_dir |
/opt/oai-cn5g/oai-cn5g-fed/docker-compose |
|
duranta_core_compose |
docker-compose-basic-nrf.yaml |
|
duranta_bringup |
benetel-bringup.sh |
canonical: RU mode+TDD, VFs, IRQ pinning, hugepages, PTP, core, DU |
duranta_gnb_cfg |
/root/oru/gnb-benetel.conf |
the image's conf; the run edits its own copy (AMF address, channel) |
duranta_du_log |
du_benetel.log |
|
duranta_band |
78 |
n78, 100 MHz (273 PRB @ 30 kHz); the centre comes from the O-RU reservation. |
duranta_center_freq_mhz |
reservation |
Channel (2026-10-16): the O-RU RESERVATION's (omf-shared-scheduler: band, centre, bandwidth), so the cell is on the spectrum you booked. |
duranta_bw_mhz |
reservation |
|
duranta_pci |
2 |
|
duranta_liteon_bringup |
liteon-du-bringup.sh |
Proven end to end 2026-09-02 .. 2026-10-25 (w43 gate on rrail: TCP DL 206-213 / UL 43). |
duranta_liteon_gnb_cfg |
/root/oru/gnb-liteon.conf |
|
duranta_liteon_du_log |
du_liteon.log |
|
duranta_liteon_center_freq_mhz |
reservation |
|
duranta_liteon_bw_mhz |
reservation |
|
duranta_liteon_pci |
1 |
set in gnb-liteon.conf (physCellId 1, nr_cellid 1) |
duranta_plmn |
00101 |
|
duranta_imsi |
auto |
auto = read from the modem's SIM (AT+CIMI); a number here is only a cross-check |
duranta_dnn |
oai |
|
duranta_ue_pool |
192.168.100.0/22 |
|
duranta_ue_mode |
modem |
modem (default): steps 4-6 register the Quectel on the second node and measure. none: the cell only - the bring-up stops once the cell is in service and touches no modem; one node (the O-DU) is enough. |
duranta_lteue_prep |
False |
true: prepare the cell and core for Amarisoft lteue (host/lteue-prep.sh): the AMF lists NIA2 first (lteue rejects the plain Security Mode Command NIA0-first gives; |
duranta_at_port |
/dev/ttyUSB2 |
|
duranta_qmi_dev |
auto |
auto = the Quectel's own QMI device and netdev, found from /dev/quectel-qmi or /dev/quectel-at (a UE node can carry two modems, e.g. a SIMCom on cdc-wdm0/wwan0 and the Quectel on cdc-wdm1/wwan1). |
duranta_wwan |
auto |
|
duranta_nr5g_bands |
77:78 |
the mask MUST include the cell's band |
duranta_remote_target |
8.8.8.8 |
|
duranta_ping_count |
10 |
|
duranta_iperf_secs |
10 |
|
duranta_iperf_port |
5201 |
|
duranta_du_timeout_s |
180 |
bring-up: RU M-Plane push + PTP settle + cell in service |
duranta_reg_timeout_s |
240 |
To change a parameter, put only the keys you change in a file of your own and pass it to any mode. A key the tutorial does not have, or a value of the wrong type, is refused before anything starts:
cat > my.yml <<'EOF'
duranta_iperf_secs: 20
EOF
cosmos-tutorial run duranta-72 --params my.yml
On a server whose fronthaul NIC has no COSMOS role name, set duranta_fh_iface to its E810 port. To change the tutorial itself, cosmos-tutorial copy duranta-72 puts an editable copy in the current directory, which then takes precedence over the installed one.
is in service while the O-RU transmits nothing useful. The evidence that the split works end to end is a DU that stays alive, fronthaul counters that climb with zero errors, and a modem that registers. This is why the parity check requires zero fronthaul errors.mimo_mode 1_2_3_4_4x4, static_compressed and DDDSU. With another DU's settings the cell carries a few Mbit/s instead of about 250, while the fronthaul stays clean and every counter looks healthy. Gate [0e] checks the stored configuration before the DU starts; gate [7] checks the whole state after the DU's M-Plane commit, because the DU itself writes registers.DDDSU cell gives roughly twice the downlink of the 40 MHz cell in step 2; the uplink stays bounded by the DL-heavy TDD pattern. Compare runs only on the same CPU platform and O-RU. Every Duranta build's measurements, by CPU platform and O-RU, are on Duranta performance by build.| Symptom | Cause | Fix |
|---|---|---|
The O-RU does not power off; omf tell replies with a reservation error |
no active O-RU reservation for this console's domain, or the run is on another domain's console | run from the console of the domain your O-RU reservation targets, while it is active |
| The bring-up exits 12 | the O-RU held another configuration and the script wrote Duranta's; it loads only at a cold boot | cold-cycle the O-RU (Step 3) and run the bring-up again, up to twice; the Ansible mode does this itself |
The bring-up stops: RU does not match the golden -- NOT starting the DU |
the O-RU's state differs from this rig's golden in a way the script does not repair | cold-cycle the O-RU and rerun; if it persists, report it to COSMOS staff with the bring-up output |
| Bring-up clean, zero fronthaul errors, but downlink a few Mbit/s with CQI 2, or no transmit power on any antenna | the O-RU's stored configuration belongs to another DU | the golden gate [0e] repairs it; cold-cycle and rerun |
The O-RU never radiates and RX_LATE is about twice RX_ON_TIME; or the DU reports RU0 mu0 (15 kHz) |
the O-RU was warm-restarted, not powered down; its management port reopened in about 80 s instead of about 300 s | repeat the cold cycle (Step 3) and confirm by ping that it went down |
The cell comes up but every UE drops out of sync (ulsch_DTX 100, BLER about 0.9) |
the O-RU was rebooted from inside instead of power-cycled; its uplink timing is wrong | a cold cycle from the console (Step 3), never an in-O-RU reboot |
| PRACH is detected but every Msg3 fails | the Benetel's own TDD pattern (its firmware carries RU-side TDD keys that the M-Plane session does not push) differs from the DU's | stage [0d] checks it; benetel-ru-tdd-check.sh fix, then cold-cycle the O-RU |
| The DU sees no PRACH at all, every other counter healthy | the O-RU is in standalone mode, so PRACH arrives on eAxC 4–7 instead of 8–9 | stage [0a] forces M-Plane mode; cold-cycle and rerun |
| The modem never decodes the cell after you changed the carrier or bandwidth | initialDLBWPcontrolResourceSetZero (CORESET#0) is per cell; a value copied from another cell or from the RFsim reference does not fit |
let the bring-up derive the cell from the reservation, or derive CORESET#0 for the new cell rather than copying it |
The bring-up aborts at its VF-MAC gate (iavf ... Reset never finished, VF disabled) |
the E810 VFs wedged after several warm DU restarts | power-cycle the O-DU server itself (omf tell -a offh, then on); a soft reboot does not clear it |
DATA1a does not exist at the start of the bring-up |
the bring-up ran before the interface naming finished after boot, or the server's fronthaul NIC has no role name | wait a minute after boot; on such a server set duranta_fh_iface |
| Fronthaul error counters climb and the PRACH noise estimate collapses | a second DU is streaming at the same O-RU | stop the other DU (the Ansible mode warns when another host holds an M-Plane session); never run two DUs against one O-RU |
nr-softmodem dies with SIGILL |
the CPU lacks AVX-512, which the image is built for | use a server that passes the image's AVX-512 check |
tcpdump on the O-DU server shows no fronthaul traffic |
DPDK takes the VF away from the kernel | use the DU's fronthaul counters, or a mirror port on the switch |
| The modem has an address but 100 % loss; TX errors rise by one per packet | the host-to-modem USB data endpoint is stuck | the helper resets the modem's USB device; see The commercial UE |
| SSH to the UE node stops answering after the data call | a default route through the modem | use only the scoped routes of Step 6 |
The Foxconn RPQN-7801 in the O-RU pool is not covered by this tutorial: with Duranta's own (CPU) physical layer it does not produce a cell a UE can see. With the GPU physical layer it does: see Duranta on NVIDIA Aerial. Its status is on O-RU hardware.
After any mode, run the teardown from the console. It stops the modem's data call and hands the modem back to ModemManager, then stops the DU, the iperf3 server and the core. The O-RU is left powered, because it takes about five minutes to come back:
OMF_NODES=<du-node>,<ue-node> cosmos-tutorial run duranta-72 --mode teardown
After the manual steps, stop the DU by the binary it is executing, never with pkill -x or pkill -f. On the O-DU server, define this helper once, then stop the DU and the core:
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
}
stop_exe /opt/duranta/cmake_targets/ran_build/build/nr-softmodem
docker compose -f /opt/oai-cn5g/oai-cn5g-fed/docker-compose/docker-compose-basic-nrf.yaml down
The bring-up launched the DU with setsid, so its process group is exactly that launch. On the UE node, stop the data call (qmicli -p -d $QMI --wds-stop-network=disable-autoconnect) and systemctl start ModemManager. Power the O-RU off only if you are done with it, from the console of your reservation's domain (omf tell -a offh -t $RU_NAME), then the nodes:
omf tell -a offh -t <du-node>,<ue-node>
Package cosmos-experiments-duranta 0.1.0-25 - on every console it is released to (cosmos-tutorial list shows the installed version). The code is installed on the console in /usr/share/cosmos-experiments/duranta-72/. To change anything, cosmos-tutorial copy duranta-72 puts an editable copy in the current directory, and cosmos-tutorial run duranta-72 then runs that copy instead of the installed one (cosmos-tutorial which duranta-72 shows which will run). Repository (GitLab account needed, not needed to run or change it): packages/cosmos-experiments-duranta/usr/share/cosmos-experiments/duranta-72.
Last verified 2026-11-02 on duranta-w44-20261101 (Duranta 2026.w44), with the tutorial sources released in cosmos-experiments-duranta 0.1.0-20, cold start (all devices off), on rrail (1× AMD EPYC 9355P, Quectel modem): three-mode parity PASS with the Benetel RAN650 (downlink 204–216, uplink 63–73 Mbit/s, ping 31 ms) and with the LiteOn FlexFi at 3500.01 MHz (downlink 185–219, uplink 46–49 Mbit/s, ping 31 ms). Earlier: three-mode parity PASS on 2026-10-30 from the installed cosmos-experiments-duranta package (rrail, Benetel, downlink 202–225 and uplink 64–75 Mbit/s), and on 2026-10-28 (sb1), 2026-10-25 and 2026-10-23. Results by build, O-RU and CPU platform are on Duranta performance by build.
References: Duranta source · Duranta at LF Networking · OAI O-RAN 7.2 fronthaul tutorial · OAI CN5G · O-RAN Alliance specifications (Open Fronthaul CUS-plane and M-plane) · linuxptp · DPDK guides · 3GPP TS 38.401 (CU and DU) · iperf3 · on this wiki: Duranta on COSMOS · O-RAN and the 7.2x fronthaul · O-RU hardware · Preparing a real-time RAN host · The commercial UE · OAI CN5G.
Author(s): COSMOS team · Package: cosmos-experiments-duranta (cosmos-tutorial run duranta-72) · Tags: duranta, oai, oran, 7.2, fronthaul, m-plane, benetel, o-ru, quectel, 5g, sa