op-packages/luci-app-sqm-autorate/docs/sqm-guide.md

SQM QoS — User Guide

luci-app-sqm-autorate is OpenWrt's standard SQM (Smart Queue Management) LuCI app, packaged for OMR with an added Autorate settings tab that wires each queue into sqm-autorate — a userspace daemon that continuously re-measures a link's real achievable bandwidth (by pinging reflectors and watching for bufferbloat) and pushes updated cake rate limits, instead of relying on the fixed Base download/upload speed you'd otherwise have to guess. It's a single page — SQM QoS under Network — with one instance per interface you want shaped.

Screenshots were taken on a test router (v0.64-snapshot) with four queue instances already defined: wan1/wan2/wan3 (plain, disabled) and omrvpn (carries real autorate/reflector tuning from prior testing — delay_thr_ms=300.0, reflector_ping_interval_s=0.5). All four are currently disabled; the Enable SQM autorate checkbox on omrvpn was turned on just for the Autorate-tab screenshot below, then reverted — it does not reflect this bench's normal running config.

https://<router-ip>/cgi-bin/luci/admin/network/sqm

If /var/run/sqm/available_qdiscs comes back empty (the sqm init script has never run), the page replaces the form with a notice and an Enable SQM button instead — not pictured here since this bench already has qdiscs available (cake, codel, fq_codel, sfq).

Each queue is a TypedSection instance (add more with the Add button at the bottom, remove one with its own Delete), with 3 tabs always present plus a 4th that only appears once autorate is turned on for that instance:

Basic Settings

Field	Meaning
Enable this SQM instance	Per-queue on/off. Checking it on any instance auto-enables the sqm init script itself (with a one-time notification reminding you to disable it manually in System Startup if that wasn't intended).
Enable SQM autorate	Turns this queue over to sqm-autorate instead of the fixed Base speeds — reveals the Autorate settings tab (see below) and makes Minimum/Maximum download/upload speed required.
Interface name	The underlying device (e.g. eth1, tun0) — a live device picker, not the OMR interface alias.
Base download/upload speed (kbit/s)	The fixed shaping rate when autorate is off; the starting rate autorate adjusts from when it's on.
Minimum/Maximum download/upload speed (autorate only)	The floor/ceiling autorate is allowed to move the rate between.
Debug logging / Verbosity	Per-instance log file under /var/run/sqm/${interface}.[start|stop]-sqm.log, and syslog verbosity (silent → trace).

A validation gotcha caught live: ticking Enable SQM autorate immediately makes the four Min/Max speed fields required — if they were never filled in (as on this bench's omrvpn, which normally runs with autorate off), they show up red/invalid right away, as in the screenshot above. Fill in all four before Save & Apply once autorate is on.

Queue Discipline

Field	Meaning
Queuing disciplines useable on this system	Populated from /var/run/sqm/available_qdiscs; must be cake if autorate is used. Installing a new qdisc kernel module needs a router restart before it shows up here.
Queue setup script	Which /usr/lib/sqm/*.qos script builds the actual tc/qdisc hierarchy — each option's help text (read from that script's .help file) is shown live under the dropdown, as in the screenshot: piece_of_cake (simplest cake-only setup, this bench's default), layer_cake (cake + diffserv layering), simple/simplest/simplest_tbf (non-cake, HTB/TBF-based alternatives).

Checking Show and Use Advanced Configuration reveals more fields — DSCP squashing/ignoring and ECN behavior on ingress/egress:

A further Show and Use Dangerous Configuration checkbox (bottom of that screenshot) reveals raw hard limits and latency targets (ilimit/elimit, itarget/etarget) plus free-form iqdisc_opts/eqdisc_opts strings passed straight to the qdisc with no validation — this is where this bench's wan1/wan2/wan3 actually carry real tuning already (iqdisc_opts: autorate-ingress dual-dsthost, eqdisc_opts: dual-srchost), enabling cake's dual-hashing so autorate's ingress shaping and per-host fairness cooperate — not shown expanded here since both checkboxes reset on every page load and only one could be captured cleanly per screenshot.

Link Layer Adaptation

Field	Meaning
Which link layer to account for	none (default — plain Ethernet WAN, this bench's setting), ethernet (VDSL2-style overhead accounting), or atm (ADSL1/2/2+ cell overhead).
Per Packet Overhead (byte) (ethernet/atm only)	Bytes to add per packet before rate-limiting, to account for lower-layer framing your ISP's link adds.
Show Advanced Linklayer Options (ethernet/atm only)	Reveals tcMTU/tcTSIZE/tcMPU size-table tuning and a linklayer adaptation mechanism override (cake/htb_private/tc_stab) — only relevant if your MTU exceeds 1500.

Autorate settings

Only present as a 4th tab once Enable SQM autorate is checked for that instance — LuCI omits the tab entirely (not just its fields) when every option inside it is hidden by that same dependency.

Field	Meaning
Output monitoring lines showing processing stats / cake bandwidth changes	Verbose per-cycle log lines from the autorate daemon, for tuning/debugging.
Debug	Even more verbose daemon logging.
Starlink support	Adjusts autorate's compensation for Starlink's own periodic latency spikes (dish reposition/satellite handoff).
Reflector ping interval in seconds	How often autorate pings its reflector set to sample latency (0.5s here).
Pingers numbers	How many concurrent reflectors to ping each cycle (4 here).
delay threshold in ms	Latency increase over baseline that autorate treats as bufferbloat and reacts to (300.0 here) — must be written as a float string (300.0, not 300); the field has no datatype validation to catch a plain integer, but cake-autorate.sh fails to start on one.
Sleep functionality	Let the daemon idle its pingers when the link itself is idle (off on this bench).
Threshold in Kbit/s below which dl/ul is considered idle	The idle-detection cutoff feeding the Sleep functionality above (500 here).
Time threshold to put pingers to sleep on sustained idle	How long the link has to stay under that threshold before pingers actually sleep (60s here).
Number of seconds to wait on startup	Delay before autorate starts adjusting rates after boot/service start, to let the link settle first (60.0s here).