Traffic control HFSC scheduler

Forward

Here I am listing what I find on this scheduler that is more complex than the HTB scheduler, but has some very good uses. It's do complex that there isn't even a man page for it currently.

There are however some references to HFSC[1] and even a post from the creater also[2], but it is still cryptic, and I'll add my findings here as I figure it out.

I'll list what I use for the rules now but I am not certan on all the values and exafctly how they all work but in time i'll fill that in

Lower the transmit queue buffer to reduce latency

ip link set txqueuelen 16 $EXT_IF

Set the MTU to 1500, although lower is better latency but worse for bulk traffic

ip link set mtu 1500 $EXT_IF

Add queue discipline with default class 50

tc qdisc add dev $EXT_IF root handle 1:0 hfsc default 50

Add root class

tc class add dev $EXT_IF parent 1:0 classid 1:1 hfsc sc rate 100kbit ul rate 100kbit

Add leaf classes

Now the fun begins.

tc class add dev $EXT_IF parent 1:1 classid 1:10 hfsc rt umax 1492b dmax 119ms rate 20kbit ls umax 1492b dmax 122ms rate 50kbit ul rate 100kbit
tc class add dev $EXT_IF parent 1:10 classid 1:2 htb rate 20kbit ceil 100kbit quantum 1500 burst 8k cburst 4k
tc class add dev $EXT_IF parent 1:10 classid 1:3 htb rate 20kbit ceil 100kbit quantum 1500 burst 6k cburst 3k
tc class add dev $EXT_IF parent 1:10 classid 1:4 htb rate 20kbit ceil 100kbit quantum 1500 burst 4k cburst 2k
tc class add dev $EXT_IF parent 1:10 classid 1:5 htb rate 20kbit ceil 100kbit quantum 1500 burst 2k cburst 1k
tc class add dev $EXT_IF parent 1:10 classid 1:6 htb rate 20kbit ceil 100kbit quantum 1500 burst 0k cburst 0k
tc class add dev $EXT_IF parent 1:1 classid 1:20 hfsc rt umax 1492b dmax 122ms rate 30kbit ls umax 1492b dmax 125ms rate 50kbit ul rate 100kbit
tc class add dev $EXT_IF parent 1:20 classid 1:21 htb rate 50kbit ceil 100kbit quantum 1500 burst 2k cburst 1k
tc class add dev $EXT_IF parent 1:20 classid 1:22 htb rate 50kbit ceil 100kbit quantum 1500 burst 0k cburst 0k
tc class add dev $EXT_IF parent 1:1 classid 1:30 hfsc rt umax 1492b dmax 125ms rate 30kbit ls umax 1492b dmax 128ms rate 50kbit ul rate 100kbit
tc class add dev $EXT_IF parent 1:1 classid 1:40 hfsc rt umax 1492b dmax 128ms rate 11kbit ls umax 1492b dmax 131ms rate 50kbit ul rate 100kbit
tc class add dev $EXT_IF parent 1:1 classid 1:50 hfsc rt umax 500b dmax 131ms rate 11kbit ls umax 1492b dmax 134ms rate 50kbit ul rate 100kbit

Attach qdisc to leaf classes

This should be explained well enough on the net.

tc qdisc add dev $EXT_IF parent 1:2 pfifo
tc qdisc add dev $EXT_IF parent 1:3 pfifo
tc qdisc add dev $EXT_IF parent 1:4 pfifo
tc qdisc add dev $EXT_IF parent 1:5 pfifo
tc qdisc add dev $EXT_IF parent 1:6 sfq perturb 10 quantum 1500
tc qdisc add dev $EXT_IF parent 1:21 sfq perturb 10 quantum 1500
tc qdisc add dev $EXT_IF parent 1:22 sfq perturb 10 quantum 1500
tc qdisc add dev $EXT_IF parent 1:30 sfq perturb 10 quantum 1500
tc qdisc add dev $EXT_IF parent 1:40 sfq perturb 10 quantum 1500
tc qdisc add dev $EXT_IF parent 1:50 sfq perturb 10 quantum 1500

Filter traffic into classes by firewall mark

tc filter add dev $EXT_IF parent 1:0 prio 0 protocol ip handle 1 fw classid 1:2
tc filter add dev $EXT_IF parent 1:0 prio 1 protocol ip handle 2 fw classid 1:3
tc filter add dev $EXT_IF parent 1:0 prio 2 protocol ip handle 3 fw classid 1:4
tc filter add dev $EXT_IF parent 1:0 prio 3 protocol ip handle 4 fw classid 1:5
tc filter add dev $EXT_IF parent 1:0 prio 4 protocol ip handle 5 fw classid 1:6
tc filter add dev $EXT_IF parent 1:0 prio 0 protocol ip handle 21 fw classid 1:21
tc filter add dev $EXT_IF parent 1:0 prio 1 protocol ip handle 22 fw classid 1:22
tc filter add dev $EXT_IF parent 1:0 prio 0 protocol ip handle 30 fw classid 1:30
tc filter add dev $EXT_IF parent 1:0 prio 0 protocol ip handle 40 fw classid 1:40
tc filter add dev $EXT_IF parent 1:0 prio 0 protocol ip handle 50 fw classid 1:50

The rest is just adding the chain and marking packets.

Fragamented thinking ahead

I'll sort this mess out once I have gathered enough information.

HFSC has an advantage over HTB but is much more complex, HFSC will alter the pirority queues depeing on virtual time delay and being there is two points on the graph, m1 and m2, you can specify a lower virtual time for m1 and a higher virtual time for m2. Why you may ask, to save time and reduce latency for interactive data while not drowning out the link on bulk data.

Command options usage:.

tc ... hfsc [[ rt SC ] [ ls SC ] | [ sc SC ]] [ ul SC ]

Before I explian all the options, most would know the brackets and what the pipe means but I will explain anyways.
You can either use rt and ls or sc but not both. ul is optional.

Note: Setting sc is the same as setting rt and ls to the same values, but as i understand it won't give any real-time value, but thinking more it's a straight line on m1 and m2 anyways so it that maybe why there is no real-time conponent to service-curve. This later sentance needs clarifying.

What the terms stand for. sc = service-curve, rt = real-time, ls = link-sharing, ul = upper-limit

Note: do not confuse sc with SC they are not the same thing

SC := [ [ m1 BPS ] [ d SEC ] m2 BPS
m1 : slope of first segment
d : x-coordinate of intersection
m2 : slope of second segment

The delay/rate value is calculated as:

8 * umax / rate = dmax

I transposed it around to also read.

(8 * umax) * dmax = umax

A example.

8 * 1500 / 0.03 = 400,000 bits per a second.
References
  1. ^ The Guide to IP Layer Network Administration with Linux was conceived as a project to document and provide examples for the use of Linux-specific IP networking tools. link. Retrieved on 2008-06-13.
  2. ^ Mail archive - turns your mailing list into a searchable archive link. Retrieved on 2008-06-13.