This is the documentation for the UCI configuration file /etc/config/qos. It is used by the package qos-scripts only.

Do NOT install multiple QoS-packages simultaneously! Uninstall the old package before installing a new one. There are at least two other QoS/ToS packages in the OpenWrt repositories regarding: dsl-qos-queue and wshaper. They do NOT use this file. dsl-qos-queue does factor in the ATM (Asynchronous Transfer Mode) overhead, which can be significant. For help writing your own script please see Traffic Control on OpenWrt: configuring the Linux Network Scheduler.

You can browse the scripts here: qos-scripts There is direct LuCI-support for qos-scripts called: luci-app-qos. NOTE: luci-app-qos won't start until you enable the qos Initscript within the System–>Startup tab as well as enable qos under Network–>QoS

As of r25641 qos-scripts dropped the use of IMQ (package iptables-mod-imq – Intermediate Queueing Device). Its successor is IFB (Intermediate Functional Block device), (requires package: kmod-ifb and the scheduler action act_connmark included).

As of r31759 qos-scripts replaced sfq/red with fq_codel to massively improve latency under load

A minimal QoS configuration usually consists of:

• one interface section
• some rules allocating packets to at least two buckets
• configuration of the buckets.

Each Interface can have its own buffer. The interface section declares global characteristics of the connection on which the specified interface is communicating. The following options are defined within this section:

config interface dsl
        option enabled      1
        option classgroup  "Default"
        option overhead     1
        option upload       512
        option download     4096

Name	Type	Required	Default	Description
enabled	boolean	yes	1	Enable/Disable QoS
classgroup	string	yes?	(none)	no idea
overhead	boolean	yes	1	decrease upload and download ratio to prevent link saturation
download	integer	yes	4096	in kilobits/second (only possible for tcp)
upload	integer	yes	512	in kilobits/second

Each classify section defines one group of packets and which target (i.e. bucket) this group belongs to. All the packets share the bucket specified.

Name	Type	Required	Default	Description
target	bucket	yes	(none)	The four defaults are: Priority, Express, Normal and Bulk
proto	string	no	0	Packets matching this protocol belong to the bucket defined in target
srchost	string	no	(none)	Packets matching this source host(s) (single IP or in CIDR notation) belong to the bucket defined in target
dsthost	string	no	(none)	Packets matching this destination host(s) (single IP or in CIDR notation) belong to the bucket defined in target
ports	integer	no	(none)	Packets matching this, belong to the bucket defined in target
srcports	integer	no	(none)	Packets matching this, belong to the bucket defined in target
dstports	integer	no	(none)	Packets matching this, belong to the bucket defined in target
portrange	integer	no	(none)	Packets matching this, belong to the bucket defined in target
pktsize	integer	no	(none)	Packets matching this, belong to the bucket defined in target
tcpflags	string	no	(none)	Packets matching this, belong to the bucket defined in target
mark	string	no	(none)	Packets matching this, belong to the bucket defined in target
layer7	string	no	(none)	Packets matching this, belong to the bucket defined in target
connbytes	int	no	(none)	Packets matching this, belong to the bucket defined in target
tos	string	no	(none)	Packets matching this, belong to the bucket defined in target
dscp	string	no	(none)	Packets matching this, belong to the bucket defined in target
direction	string	no	(none)	Packets matching this traffic direction (in or out) belong to the bucket defined in target

As we can have more then one interface, we can have more then one classgroup.

Name	Type	Required	Default	Description
classes	bucket names	yes	(none)	Specifies the number of buckets and their names
default	bucket name	yes	(none)	Defines which bucket is considered default

Each Bucket has its own configuration.

Name	Type	Required	Default	Description
packetsize	integer	yes	(none)	in bytes
packetdelay	integer	yes	(none)	in ms
maxsize	integer	yes	(none)	in bytes
avgrate	integer	yes	(none)	unknown, value in %
limitrate	integer	no	100	Defines to how much percent of the available bandwidth this class is capped to, value in %
maxsize	integer	yes	(none)	in bytes
priority	integer	yes	(none)	in %

1. Install the qos-scripts package:

opkg install qos-scripts

2. Basic configuration using UCI command line:

uci set qos.wan.upload=1000            # Upload speed in kBits/s
uci set qos.wan.download=16000         # Download speed in kBits/s
uci set qos.wan.enabled=1
uci commit qos

3. Start it and look for error output and test):

/etc/init.d/qos start

4. Make script run at every boot up:

/etc/init.d/qos enable

Recent versions of trunk uses CoDel (pronounced: Coddle), so this should not be needed. bufferbloat.net

( Note: after i know about bufferbloat - websearchd and many misguided users are asking about raising txqueuelen, or how to set it, and this was a recommended read posted for them -, and that just enabling QoS and setting up the rules i wanted didnt produce results, i thought of reduce txqueuelen from the default size of 1000 packets. On most SOHO applications the upload speed is much-much lower than the interface speed (100M or 1G), and it's written that the def buffer is tailored for enterprise size usage. I used values of 90 where[only] it was 1000 and it is wonderfully responsive and effective now, on a 256k connection. I also raised lenghts of 3 and 5 to 20 as i imagined it might be hard to do any queuing/shaping on such ultra short buffers. Leave the 0-s as 0. Btw, the 90 came by, that with 1500byte packet size and 256kbit speed, it takes just half a sec to empty it. Well, i mistakenly took mtu as bits so it's 4 second really but i didnt have the courage for a drastically smaller than default value, 90 already seem so smallish, and it works nicely anyway, awesome lack of latency, jitter and packetloss on the other machine, according to http://pingtest.net, before of this there was no difference between mine and that. Note that this time doesnt correspond to ping values. The point is to allow built-in TCP congestion control to work to reduce "spamming", set speeds as it was envisioned, and keep things leveled out rather than fluctuating widely. While this doesnt directly effects QoS-ing itself, it is extremely beneficial, even essential, to the results usually expected from employing it. It sets a foundation, a healthy network environment over which QoS to function.

First we list the stuff with

ifconfig

. Then apply for example

ifconfig eth0 txqueuelen 90

and

ifconfig pppoe-wan txqueuelen 20

etc accordingly. Finally an

uci commit
/etc/init.d/network reload

might be needed, i did a network restart too after this.)

The qos-scripts package didn't come with documentation and there has been some confusion about its features, among users. The information in this section comes straight from nbd (the developer), so it should come a long way to clearing some confusion on two major issues.

The biggest item of contention was which group setting gives better performance, Priority or Express. As it turns out, it depends on the application. Priority boosts low-bandwidth small frames, such as TCP-ACKs and DNS more than Express. Express is for prioritizing bigger frames, which would include stuff like VoIP (port 5060).

Another biggie was the exact meaning of each type. Types are necessary for connection tracking. By default, Classify is not run on a connection that had already been assigned a traffic class, so it is the initial connection-tracked classification. Reclassify can override the traffic class per packet, without altering the connection tracking mark. Default is a fall-back for everything that has not been marked by Classify/Reclassify. Rules get processed by type first (Classify gets processed first, then Reclassify and finally Default) and then based on the order in the configuration file (top to bottom).