Donate

Categories

Advert

XHTML

Valid XHTML 1.0 Transitional

Linux Rate-Limiting of an ADSL Link

After great pain I’ve got tc working on some Linux routers. The difficulty with limiting an ADSL link is that the ADSL modem has significant buffers and the link between the Linux machine and the modem is significantly faster than the ADSL upstream channel. This means that the transmission speed needs to be artificially limited, a speed of about 95% the maximum channel speed is often recommended. As ADSL upstream speed often varies (at least in my experience) that means that you must limit the transmission speed to 95% of the lowest speed that you expect to see – which of course means a significant drop in performance when the ADSL link is performing well.

I use the HTB queuing discipline to limit the transmission rate. My transmission speed varies between 550kbit and 680kbit in my rough tests. So I start by limiting the overall device performance to 550kbit. Then I have three different classes with IDs 1:10, 1:20, and 1:30 with rates of 64kbit, 480kbit, and 128kbit respectively. It is often recommended that the inferior classes have a total bandwidth allowance that is equal to the allowance for the overall link, but I have the three inferior classes allocated with 672kbit – I think that this will work as it will be quite rare that all classes will be in operation at the same time and fairly unlikely that I will ever have all three classes running at maximum speed. I will be interested to see any comments about this, I might have misunderstood the issues related to this.

Each class has a SFQ queue discipline associated with it for fair queuing within the class. It might be a bit of overkill, I expect to only have one data channel in operation on the VOIP class so it probably does no good there and my usage pattern is such that if the 480kbit connection is anywhere near busy then it’s due to a single large transfer. But with the power of a P3 CPU applied to the task of routing at ADSL speeds it really doesn’t matter if some CPU time is wasted.

Then the tc filter lines associate iptables marks with the classes.

Now this is only a tiny fraction of what tc can do. But I think that this basic configuration with the rate limits changed will suit many ADSL router configurations, it may not be an ideal configuration for most ADSL routers but it will probably be a viable configuration that will be better than having no traffic shaping. Below is the shell script that I am using:

#!/bin/bash -e

DEV=ppp0

tc qdisc del dev $DEV parent root handle 1:0 2> /dev/null | true
tc qdisc add dev $DEV parent root handle 1:0 htb default 30

# limit the rate to slightly lower than DSL line speed
tc class add dev $DEV parent 1:0 classid 1:1 htb rate 550kbit prio 1

# sub classes for each traffic type
# 10 is VOIP, 20 is default, 30 is the test network
tc class add dev $DEV parent 1:1 classid 1:10 htb rate 64kbit burst 6k prio 2
tc class add dev $DEV parent 1:1 classid 1:20 htb rate 480kbit burst 12k prio 3
tc class add dev $DEV parent 1:1 classid 1:30 htb rate 128kbit burst 12k prio 4

# use an sfq under each class to share the bandwidth
tc qdisc add dev $DEV parent 1:10 handle 10: sfq
tc qdisc add dev $DEV parent 1:20 handle 20: sfq
tc qdisc add dev $DEV parent 1:30 handle 30: sfq

tc filter add dev $DEV parent 1: protocol ip prio 1 handle 1 fw classid 1:10
tc filter add dev $DEV parent 1: protocol ip prio 2 handle 2 fw classid 1:20
tc filter add dev $DEV parent 1: protocol ip prio 3 handle 3 fw classid 1:30

iptables -t mangle -F POSTROUTING
iptables -t mangle -A POSTROUTING -j MARK --set-mark 2
iptables -t mangle -A POSTROUTING -p tcp --sport 22 -j MARK --set-mark 3
iptables -t mangle -A POSTROUTING -d $VOIPSERVER -j MARK --set-mark 1

2 comments to Linux Rate-Limiting of an ADSL Link

  • Why not just:

    apt-get install wondershaper

    See also http://lartc.org/wondershaper and the rest of the lartc site of course.

  • Glen Turner

    Commonly a burst size less than a TCP segment size results in the same performance in the TCP traffic classes as having no burst configured at all.

    Continuing that thought, is the reduction in TCP throughput from having a burst size less than the bandwidth-delay product acceptable for a bottleneck link? Or will the result be to have the traffic class’s buffering act as the path’s bottleneck?