Skip to content

power-steering.md

Latest commit

 

History

History
90 lines (68 loc) · 7.16 KB

power-steering.md

File metadata and controls

90 lines (68 loc) · 7.16 KB
Raw

5 GHz Tx power steering on wlc-01

In a dense session hall, RRM leaves a few radios with ~100 clients (the per-radio cap) while neighbours idle. Lowering the hot radios' Tx power shrinks their cells so clients pick neighbours. This was done by hand at Kaigi on Rails 2026 and worked; the steps are in ../events/kor26/power-steering.md. The procedure Claude sessions follow is the rknet-wlc-power-steering skill.

What worked and what did not

  • Static power on a radio at the cap, one step (3 dB) at a time, moved 28–46 clients to neighbours within seconds, and cap rejects stopped. A radio that refilled took one more step (14 → 11 dBm).
  • The neighbour that absorbs the load can reach the cap within a minute and becomes the next step (trb-12 → trb-13, lb1-01 → lb1-31).
  • An RF profile tx-power-max cap (17 dBm) lowered the hall's radios, but hot radios refilled within ten minutes as the crowd grew. A cap alone does not relieve them.
  • Load balancing did not fix distribution (below).
  • Raising max-clients was rejected: a 20 MHz radio cannot serve 200 clients well, and it does nothing for distribution.
  • Failed-frame ratio is not a reason to steer. lb1-03 was cut 17 → 14 dBm as a single test: its ratio went 24.9% → 16.1% → 23.7% and untouched neighbours got worse. The ratio follows traffic volume and constant re-association (management and short-frame retries), not cell size.
  • Short spikes to 90–100 often fell back on their own within minutes. Act on a radio that stays at the cap or keeps producing rejects.

Reading load

Signal Where
Clients per radio show client summary counted per AP for slot 1 (fresh); bsnApIfNoOfUsers{bsnAPIfSlotId="1"} in Prometheus (lags)
Cap rejections syslog %APF-4-ASSOCREQ_PROC_FAILED ... Maximum Client Limit Reached in AP Radio. Reject new Association. on wlc-01; the line has no AP name
Power and channel per radio show advanced 802.11a summary (slot column) and show advanced 802.11a txpower
Neighbours and client RSSI/SNR show ap auto-rf 802.11a <ap>
  • Count slot 1 only. Per-AP totals include 2.4 GHz.
  • Ghost entries (Association Id 0) show as Associated and inflate the WLC count (troubleshooting.md).
  • SNMP lags. Right after a radio cycle Prometheus showed a 140-client "drop" while the WLC already had everyone back, and it showed 200/200 on a radio the WLC listed at 172.
  • Failure ratios on idle radios, or during a burst of new joins, mean nothing. WLC channel utilization (0–1%) was not meaningful under load.

Why RRM leaves dense radios at max power

  • TPC runs TPCv2 (target RSSI -67 dBm, path-loss model). It is coverage-driven and pushed the track A radios to their hardware max (23 dBm) although each heard 8–22 neighbours at -30 to -57 dBm.
  • TPCv1 instead lowers power until the third-loudest neighbour hears the AP at the threshold (-65 dBm in rk-5g-hi). With neighbours at -47 to -50 it would drop those radios to about 8 dBm, likely too far with a crowd. Consider it off-peak, possibly with a gentler threshold.
  • TPC moves in steps every 600 s (TPCv2 Target Plan: To be reached in N TPC runs).
  • AireOS load balancing only denies at association time and never moves associated clients. Enabled on WLAN 10, it denied 60–100 clients a minute without fixing distribution.
  • The per-radio client cap is not load balancing. Clients rejected by the cap mostly took minutes (median about 8 min) to get on, often onto the same hot radios.

Where settings live: RF profiles override global

Every AP is in an AP group with a 5 GHz RF profile. The profile's TPC min/max, max clients, rates and load-balancing window/denial override the global 802.11a and load-balancing settings, so config 802.11a txPower global max has no effect.

AP group 5 GHz RF profile APs (Kaigi on Rails 2026)
rk-hi rk-5g-hi tra, trb, lb1, fo0, noc
rk-med rk-5g-med lb0, int, off 
show wlan apgroups                    group membership and profiles
show rf-profile details rk-5g-hi      Min/Max Transmit Power, Max Clients, TPC thresholds
show advanced 802.11a txpower         Last Run, plan status, per-radio power
  • Kaigi on Rails 2026 values: rk-5g-hi min 7 / max 30 dBm, Max Clients 100, 20 MHz, low rates disabled; rk-5g-med min -10 / max 30, Max Clients 200. rk-med radios ran 19–22 dBm, so one step there is bigger.
  • The profile minimum is not honoured on XOR slot-0 radios of 2802 APs. They sit at 2 dBm.
  • Channels are 20 MHz, so 36/40/44/48 do not overlap. Neighbours on adjacent channels at -50 dBm still add some interference.

Lever 1: RF profile power limits (non-disruptive)

config rf-profile tx-power-max 17 rk-5g-hi      # rollback: ... tx-power-max 30 rk-5g-hi
config rf-profile tx-power-min 14 rk-5g-hi      # rollback: ... tx-power-min 7 rk-5g-hi
config advanced 802.11a txpower-update          # run TPC now without changing its mode
  • Profile edits drop no client and take effect at the next TPC run.
  • config 802.11a txPower global once also triggers a run, but it may switch TPC to one-shot mode. Prefer txpower-update.
  • Some underused neighbours sat at 8 dBm under TPC and showed the worst failure ratios. Raising the profile minimum counters that.

Lever 2: per-AP static power (drops that radio's clients)

config 802.11a txPower ap <ap> <level> acts on slot 1. While the radio is up it is refused with Radio [Slot 1] is not Disabled or Customize is not enabled, and no CLI "customize" switch was found. The working sequence:

config 802.11a disable <ap>
config 802.11a txPower ap <ap> <level>       # or `global` to return it to TPC
config 802.11a enable <ap>
  • The radio is down 1–2 s. Its clients re-associate within seconds, a good share onto neighbours.
  • show ap config 802.11a <ap> then shows Tx Power Configuration ... CUSTOMIZED, and TPC no longer manages that radio.
  • Levels index the radio's own power table (ap-x CH* L/N (X dBm) [table] in show advanced 802.11a txpower). A * before the level means TPC manages it; no * means static. XOR APs list two 5 GHz radios there without a slot column, so read slot 1 from show advanced 802.11a summary (columns AP, MAC, slot, admin, oper, channel, level). The table depends on AP model and channel:
    • [23/20/17/14/11/8/5/2]: level 3 = 17 dBm, level 4 = 14 dBm
    • [22/19/16/13/10/7/4/2] (AP1852I and rk-med radios)
    • [17/14/11/8/5/2]: level 1 = 17 dBm, level 2 = 14 dBm
    • [16/13/10/7/4/2]
  • A static setting stores the level index, not dBm. When DCA or a reboot moves the radio to another channel its dBm changes (trb-11 level 4 went from 14 to 8 dBm overnight; ap-off-01 level 3 from 16 to 10). Re-read every CUSTOMIZED radio after a channel change, an AP reboot and at the start of each day.
  • Static levels survive AP reboots and mode changes, and stay after the event until someone returns them to global.

Rollback

Once the crowd thins, return CUSTOMIZED radios to global and the profile to its previous limits, one AP at a time at a break. At Kaigi on Rails 2026 this was not done and the settings stayed on the WLC (../events/kor26/README.md).