10.4. Quality of Service

10.4.1. Principle and Mechanism

Quality of Service (or QoS for short) refers to a set of techniques that guarantee or improve the quality of the service provided to applications. The most popular such technique involves classifying the network traffic into categories, and differentiating the handling of traffic according to which category it belongs to. The main application of this differentiated services concept is traffic shaping, which limits the data transmission rates for connections related to some services and/or hosts so as not to saturate the available bandwidth and starve important other services. Traffic shaping is a particularly good fit for TCP traffic, since this protocol automatically adapts to available bandwidth.

It is also possible to alter the priorities on traffic, which allows prioritizing packets related to interactive services (such as ssh and telnet) or to services that only deal with small blocks of data.

The Debian kernels include the features required for QoS along with their associated modules. These modules are many, and each of them provides a different service, most notably by way of special schedulers for the queues of IP packets; the wide range of available scheduler behaviors spans the whole range of possible requirements.

10.4.2. Configuring and Implementing

QoS parameters are set through the tc command (provided by the iproute package). Since its interface is quite complex, using higher-level tools is recommended.

10.4.2.1. Reducing Latencies: `wondershaper`

The main purpose of wondershaper (in the similarly-named package) is to minimize latencies independent of network load. This is achieved by limiting total traffic to a value that falls just short of the link saturation value.

Once a network interface is configured, setting up this traffic limitation is achieved by running wondershaper interface download_rate upload_rate. The interface can be enp1s0, eth0 or ppp0 for example, and both rates are expressed in kilobits per second. The wondershaper remove interface command disables traffic control on the specified interface.

For an Ethernet connection, historically this script would be called right after the interface is configured. This is done by adding up and down directives to the /etc/network/interfaces file allowing declared commands to be run, respectively, after the interface is configured and before it is deconfigured. Or in the PPP case, creating a script that calls wondershaper in /etc/ppp/ip-up.d/ will enable traffic control as soon as the connection is up. Below is an example using this first method:

Example 10.9. Changes in the /etc/network/interfaces file

iface eth0 inet dhcp
    up /sbin/wondershaper eth0 500 100
    down /sbin/wondershaper remove eth0

10.4.2.2. Standard Configuration

Barring a specific QoS configuration, the Linux kernel uses the pfifo_fast queue scheduler, which provides a few interesting features by itself. The priority of each processed IP packet is based on the DSCP field (Differentiated Services Code Point) of this packet; modifying this 6-bit field is enough to take advantage of the scheduling features. Refer to https://en.wikipedia.org/wiki/Differentiated_services#Class_Selector for more information.

The DSCP field can be set by applications that generate IP packets, or modified on the fly by netfilter. The following rules are sufficient to increase responsiveness for a server's SSH service, note that the DSCP field must be set in hexadecimal:

nft add table ip mangle
nft add rule ip mangle PREROUTING tcp sport 22 counter ip dscp set 0x04
nft add rule ip mangle PREROUTING tcp dport 22 counter ip dscp set 0x04