9.5. Classful Queueing Disciplines

Classful qdiscs are very useful if you have different kinds of traffic which should have differing treatment. One of the classful qdiscs is called 'CBQ', 'Class Based Queueing' and it is so widely mentioned that people identify queueing with classes solely with CBQ, but this is not the case.

CBQ is merely the oldest kid on the block - and also the most complex one. It may not always do what you want. This may come as something of a shock to many who fell for the 'sendmail effect', which teaches us that any complex technology which doesn't come with documentation must be the best available.

More about CBQ and its alternatives shortly.

9.5.1. Flow within classful qdiscs & classes

When traffic enters a classful qdisc, it needs to be sent to any of the classes within - it needs to be 'classified'. To determine what to do with a packet, the so called 'filters' are consulted. It is important to know that the filters are called from within a qdisc, and not the other way around!

The filters attached to that qdisc then return with a decision, and the qdisc uses this to enqueue the packet into one of the classes. Each subclass may try other filters to see if further instructions apply. If not, the class enqueues the packet to the qdisc it contains.

Besides containing other qdiscs, most classful qdiscs also perform shaping. This is useful to perform both packet scheduling (with SFQ, for example) and rate control. You need this in cases where you have a high speed interface (for example, ethernet) to a slower device (a cable modem).

If you were only to run SFQ, nothing would happen, as packets enter & leave your router without delay: the output interface is far faster than your actual link speed. There is no queue to schedule then.

9.5.2. The qdisc family: roots, handles, siblings and parents

Each interface has one egress 'root qdisc'. By default, it is the earlier mentioned classless pfifo_fast queueing discipline. Each qdisc and class is assigned a handle, which can be used by later configuration statements to refer to that qdisc. Besides an egress qdisc, an interface may also have an ingress qdisc , which polices traffic coming in.

The handles of these qdiscs consist of two parts, a major number and a minor number : <major>:<minor>. It is customary to name the root qdisc '1:', which is equal to '1:0'. The minor number of a qdisc is always 0.

Classes need to have the same major number as their parent. This major number must be unique within a egress or ingress setup. The minor number must be unique within a qdisc and his classes.

9.5.2.1. How filters are used to classify traffic

Recapping, a typical hierarchy might look like this:

                     1:   root qdisc
                      |
                     1:1    child class
                   /  |  \
                  /   |   \
                 /    |    \
                 /    |    \
              1:10  1:11  1:12   child classes
               |      |     | 
               |     11:    |    leaf class
               |            | 
               10:         12:   qdisc
              /   \       /   \
           10:1  10:2   12:1  12:2   leaf classes

But don't let this tree fool you! You should *not* imagine the kernel to be at the apex of the tree and the network below, that is just not the case. Packets get enqueued and dequeued at the root qdisc, which is the only thing the kernel talks to.

A packet might get classified in a chain like this:

1: -> 1:1 -> 1:12 -> 12: -> 12:2

The packet now resides in a queue in a qdisc attached to class 12:2. In this example, a filter was attached to each 'node' in the tree, each choosing a branch to take next. This can make sense. However, this is also possible:

1: -> 12:2

In this case, a filter attached to the root decided to send the packet directly to 12:2.

9.5.2.2. How packets are dequeued to the hardware

When the kernel decides that it needs to extract packets to send to the interface, the root qdisc 1: gets a dequeue request, which is passed to 1:1, which is in turn passed to 10:, 11: and 12:, each of which queries its siblings, and tries to dequeue() from them. In this case, the kernel needs to walk the entire tree, because only 12:2 contains a packet.

In short, nested classes ONLY talk to their parent qdiscs, never to an interface. Only the root qdisc gets dequeued by the kernel!

The upshot of this is that classes never get dequeued faster than their parents allow. And this is exactly what we want: this way we can have SFQ in an inner class, which doesn't do any shaping, only scheduling, and have a shaping outer qdisc, which does the shaping.

9.5.3. The PRIO qdisc

The PRIO qdisc doesn't actually shape, it only subdivides traffic based on how you configured your filters. You can consider the PRIO qdisc a kind of pfifo_fast on steroids, whereby each band is a separate class instead of a simple FIFO.

When a packet is enqueued to the PRIO qdisc, a class is chosen based on the filter commands you gave. By default, three classes are created. These classes by default contain pure FIFO qdiscs with no internal structure, but you can replace these by any qdisc you have available.

Whenever a packet needs to be dequeued, class :1 is tried first. Higher classes are only used if lower bands all did not give up a packet.

This qdisc is very useful in case you want to prioritize certain kinds of traffic without using only TOS-flags but using all the power of the tc filters. You can also add an other qdisc to the 3 predefined classes, whereas pfifo_fast is limited to simple fifo qdiscs.

Because it doesn't actually shape, the same warning as for SFQ holds: either use it only if your physical link is really full or wrap it inside a classful qdisc that does shape. The latter holds for almost all cable modems and DSL devices.

In formal words, the PRIO qdisc is a Work-Conserving scheduler.

9.5.3.1. PRIO parameters & usage

The following parameters are recognized by tc:

bands

Number of bands to create. Each band is in fact a class. If you change this number, you must also change:

priomap

If you do not provide tc filters to classify traffic, the PRIO qdisc looks at the TC_PRIO priority to decide how to enqueue traffic.

This works just like with the pfifo_fast qdisc mentioned earlier, see there for lots of detail.

The bands are classes, and are called major:1 to major:3 by default, so if your PRIO qdisc is called 12:, tc filter traffic to 12:1 to grant it more priority.

Reiterating, band 0 goes to minor number 1! Band 1 to minor number 2, etc.

9.5.3.2. Sample configuration

We will create this tree:

          1:   root qdisc
         / | \ 
       /   |   \
       /   |   \
     1:1  1:2  1:3    classes
      |    |    |
     10:  20:  30:    qdiscs    qdiscs
     sfq  tbf  sfq
band  0    1    2

Bulk traffic will go to 30:, interactive traffic to 20: or 10:.

Command lines:

# tc qdisc add dev eth0 root handle 1: prio 
## This *instantly* creates classes 1:1, 1:2, 1:3
  
# tc qdisc add dev eth0 parent 1:1 handle 10: sfq
# tc qdisc add dev eth0 parent 1:2 handle 20: tbf rate 20kbit buffer 1600 limit 3000
# tc qdisc add dev eth0 parent 1:3 handle 30: sfq                                

Now let's see what we created:

# tc -s qdisc ls dev eth0 
qdisc sfq 30: quantum 1514b 
 Sent 0 bytes 0 pkts (dropped 0, overlimits 0) 

 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms 
 Sent 0 bytes 0 pkts (dropped 0, overlimits 0) 

 qdisc sfq 10: quantum 1514b 
 Sent 132 bytes 2 pkts (dropped 0, overlimits 0) 

 qdisc prio 1: bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
 Sent 174 bytes 3 pkts (dropped 0, overlimits 0) 
As you can see, band 0 has already had some traffic, and one packet was sent while running this command!

We now do some bulk data transfer with a tool that properly sets TOS flags, and take another look:

# scp tc ahu@10.0.0.11:./
ahu@10.0.0.11's password: 
tc                   100% |*****************************|   353 KB    00:00    
# tc -s qdisc ls dev eth0
qdisc sfq 30: quantum 1514b 
 Sent 384228 bytes 274 pkts (dropped 0, overlimits 0) 

 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms 
 Sent 2640 bytes 20 pkts (dropped 0, overlimits 0) 

 qdisc sfq 10: quantum 1514b 
 Sent 2230 bytes 31 pkts (dropped 0, overlimits 0) 

 qdisc prio 1: bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
 Sent 389140 bytes 326 pkts (dropped 0, overlimits 0) 
As you can see, all traffic went to handle 30:, which is the lowest priority band, just as intended. Now to verify that interactive traffic goes to higher bands, we create some interactive traffic:

# tc -s qdisc ls dev eth0
qdisc sfq 30: quantum 1514b 
 Sent 384228 bytes 274 pkts (dropped 0, overlimits 0) 

 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms 
 Sent 2640 bytes 20 pkts (dropped 0, overlimits 0) 

 qdisc sfq 10: quantum 1514b 
 Sent 14926 bytes 193 pkts (dropped 0, overlimits 0) 

 qdisc prio 1: bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
 Sent 401836 bytes 488 pkts (dropped 0, overlimits 0) 

It worked - all additional traffic has gone to 10:, which is our highest priority qdisc. No traffic was sent to the lowest priority, which previously received our entire scp.

9.5.4. The famous CBQ qdisc

As said before, CBQ is the most complex qdisc available, the most hyped, the least understood, and probably the trickiest one to get right. This is not because the authors are evil or incompetent, far from it, it's just that the CBQ algorithm isn't all that precise and doesn't really match the way Linux works.

Besides being classful, CBQ is also a shaper and it is in that aspect that it really doesn't work very well. It should work like this. If you try to shape a 10mbit/s connection to 1mbit/s, the link should be idle 90% of the time. If it isn't, we need to throttle so that it IS idle 90% of the time.

This is pretty hard to measure, so CBQ instead derives the idle time from the number of microseconds that elapse between requests from the hardware layer for more data. Combined, this can be used to approximate how full or empty the link is.

This is rather tortuous and doesn't always arrive at proper results. For example, what if the actual link speed of an interface that is not really able to transmit the full 100mbit/s of data, perhaps because of a badly implemented driver? A PCMCIA network card will also never achieve 100mbit/s because of the way the bus is designed - again, how do we calculate the idle time?

It gets even worse if we consider not-quite-real network devices like PPP over Ethernet or PPTP over TCP/IP. The effective bandwidth in that case is probably determined by the efficiency of pipes to userspace - which is huge.

People who have done measurements discover that CBQ is not always very accurate and sometimes completely misses the mark.

In many circumstances however it works well. With the documentation provided here, you should be able to configure it to work well in most cases.

9.5.4.1. CBQ shaping in detail

As said before, CBQ works by making sure that the link is idle just long enough to bring down the real bandwidth to the configured rate. To do so, it calculates the time that should pass between average packets.

During operations, the effective idletime is measured using an exponential weighted moving average (EWMA), which considers recent packets to be exponentially more important than past ones. The UNIX loadaverage is calculated in the same way.

The calculated idle time is subtracted from the EWMA measured one, the resulting number is called 'avgidle'. A perfectly loaded link has an avgidle of zero: packets arrive exactly once every calculated interval.

An overloaded link has a negative avgidle and if it gets too negative, CBQ shuts down for a while and is then 'overlimit'.

Conversely, an idle link might amass a huge avgidle, which would then allow infinite bandwidths after a few hours of silence. To prevent this, avgidle is capped at maxidle.

If overlimit, in theory, the CBQ could throttle itself for exactly the amount of time that was calculated to pass between packets, and then pass one packet, and throttle again. But see the 'minburst' parameter below.

These are parameters you can specify in order to configure shaping:

avpkt

Average size of a packet, measured in bytes. Needed for calculating maxidle, which is derived from maxburst, which is specified in packets.

bandwidth

The physical bandwidth of your device, needed for idle time calculations.

cell

The time a packet takes to be transmitted over a device may grow in steps, based on the packet size. An 800 and an 806 size packet may take just as long to send, for example - this sets the granularity. Most often set to '8'. Must be an integral power of two.

maxburst

This number of packets is used to calculate maxidle so that when avgidle is at maxidle, this number of average packets can be burst before avgidle drops to 0. Set it higher to be more tolerant of bursts. You can't set maxidle directly, only via this parameter.

minburst

As mentioned before, CBQ needs to throttle in case of overlimit. The ideal solution is to do so for exactly the calculated idle time, and pass 1 packet. For Unix kernels, however, it is generally hard to schedule events shorter than 10ms, so it is better to throttle for a longer period, and then pass minburst packets in one go, and then sleep minburst times longer.

The time to wait is called the offtime. Higher values of minburst lead to more accurate shaping in the long term, but to bigger bursts at millisecond timescales.

minidle

If avgidle is below 0, we are overlimits and need to wait until avgidle will be big enough to send one packet. To prevent a sudden burst from shutting down the link for a prolonged period of time, avgidle is reset to minidle if it gets too low.

Minidle is specified in negative microseconds, so 10 means that avgidle is capped at -10us.

mpu

Minimum packet size - needed because even a zero size packet is padded to 64 bytes on ethernet, and so takes a certain time to transmit. CBQ needs to know this to accurately calculate the idle time.

rate

Desired rate of traffic leaving this qdisc - this is the 'speed knob'!

Internally, CBQ has a lot of fine tuning. For example, classes which are known not to have data enqueued to them aren't queried. Overlimit classes are penalized by lowering their effective priority. All very smart & complicated.

9.5.4.2. CBQ classful behaviour

Besides shaping, using the aforementioned idletime approximations, CBQ also acts like the PRIO queue in the sense that classes can have differing priorities and that lower priority numbers will be polled before the higher priority ones.

Each time a packet is requested by the hardware layer to be sent out to the network, a weighted round robin process ('WRR') starts, beginning with the lower-numbered priority classes.

These are then grouped and queried if they have data available. If so, it is returned. After a class has been allowed to dequeue a number of bytes, the next class within that priority is tried.

The following parameters control the WRR process:

allot

When the outer CBQ is asked for a packet to send out on the interface, it will try all inner qdiscs (in the classes) in turn, in order of the 'priority' parameter. Each time a class gets its turn, it can only send out a limited amount of data. 'Allot' is the base unit of this amount. See the 'weight' parameter for more information.

prio

The CBQ can also act like the PRIO device. Inner classes with higher priority are tried first and as long as they have traffic, other classes are not polled for traffic.

weight

Weight helps in the Weighted Round Robin process. Each class gets a chance to send in turn. If you have classes with significantly more bandwidth than other classes, it makes sense to allow them to send more data in one round than the others.

A CBQ adds up all weights under a class, and normalizes them, so you can use arbitrary numbers: only the ratios are important. People have been using 'rate/10' as a rule of thumb and it appears to work well. The renormalized weight is multiplied by the 'allot' parameter to determine how much data can be sent in one round.

Please note that all classes within an CBQ hierarchy need to share the same major number!

9.5.4.3. CBQ parameters that determine link sharing & borrowing

Besides purely limiting certain kinds of traffic, it is also possible to specify which classes can borrow capacity from other classes or, conversely, lend out bandwidth.

Isolated/sharing

A class that is configured with 'isolated' will not lend out bandwidth to sibling classes. Use this if you have competing or mutually-unfriendly agencies on your link who do not want to give each other freebies.

The control program tc also knows about 'sharing', which is the reverse of 'isolated'.

bounded/borrow

A class can also be 'bounded', which means that it will not try to borrow bandwidth from sibling classes. tc also knows about 'borrow', which is the reverse of 'bounded'.

A typical situation might be where you have two agencies on your link which are both 'isolated' and 'bounded', which means that they are really limited to their assigned rate, and also won't allow each other to borrow.

Within such an agency class, there might be other classes which are allowed to swap bandwidth.

9.5.4.4. Sample configuration


               1:           root qdisc
               |
              1:1           child class
             /   \
            /     \
          1:3     1:4       leaf classes
           |       |
          30:     40:       qdiscs
         (sfq)   (sfq)

This configuration limits webserver traffic to 5mbit and SMTP traffic to 3 mbit. Together, they may not get more than 6mbit. We have a 100mbit NIC and the classes may borrow bandwidth from each other.

# tc qdisc add dev eth0 root handle 1:0 cbq bandwidth 100Mbit         \
  avpkt 1000 cell 8
# tc class add dev eth0 parent 1:0 classid 1:1 cbq bandwidth 100Mbit  \
  rate 6Mbit weight 0.6Mbit prio 8 allot 1514 cell 8 maxburst 20      \
  avpkt 1000 bounded
This part installs the root and the customary 1:1 class. The 1:1 class is bounded, so the total bandwidth can't exceed 6mbit.

As said before, CBQ requires a *lot* of knobs. All parameters are explained above, however. The corresponding HTB configuration is lots simpler.

# tc class add dev eth0 parent 1:1 classid 1:3 cbq bandwidth 100Mbit  \
  rate 5Mbit weight 0.5Mbit prio 5 allot 1514 cell 8 maxburst 20      \
  avpkt 1000                       
# tc class add dev eth0 parent 1:1 classid 1:4 cbq bandwidth 100Mbit  \
  rate 3Mbit weight 0.3Mbit prio 5 allot 1514 cell 8 maxburst 20      \
  avpkt 1000

These are our two leaf classes. Note how we scale the weight with the configured rate. Both classes are not bounded, but they are connected to class 1:1 which is bounded. So the sum of bandwith of the 2 classes will never be more than 6mbit. The classids need to be within the same major number as the parent qdisc, by the way!

# tc qdisc add dev eth0 parent 1:3 handle 30: sfq
# tc qdisc add dev eth0 parent 1:4 handle 40: sfq

Both classes have a FIFO qdisc by default. But we replaced these with an SFQ queue so each flow of data is treated equally.

# tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
  sport 80 0xffff flowid 1:3
# tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
  sport 25 0xffff flowid 1:4

These commands, attached directly to the root, send traffic to the right qdiscs.

Note that we use 'tc class add' to CREATE classes within a qdisc, but that we use 'tc qdisc add' to actually add qdiscs to these classes.

You may wonder what happens to traffic that is not classified by any of the two rules. It appears that in this case, data will then be processed within 1:0, and be unlimited.

If SMTP+web together try to exceed the set limit of 6mbit/s, bandwidth will be divided according to the weight parameter, giving 5/8 of traffic to the webserver and 3/8 to the mail server.

With this configuration you can also say that webserver traffic will always get at minimum 5/8 * 6 mbit = 3.75 mbit.

9.5.4.5. Other CBQ parameters: split & defmap

As said before, a classful qdisc needs to call filters to determine which class a packet will be enqueued to.

Besides calling the filter, CBQ offers other options, defmap & split. This is pretty complicated to understand, and it is not vital. But as this is the only known place where defmap & split are properly explained, I'm doing my best.

As you will often want to filter on the Type of Service field only, a special syntax is provided. Whenever the CBQ needs to figure out where a packet needs to be enqueued, it checks if this node is a 'split node'. If so, one of the sub-qdiscs has indicated that it wishes to receive all packets with a certain configured priority, as might be derived from the TOS field, or socket options set by applications.

The packets' priority bits are and-ed with the defmap field to see if a match exists. In other words, this is a short-hand way of creating a very fast filter, which only matches certain priorities. A defmap of ff (hex) will match everything, a map of 0 nothing. A sample configuration may help make things clearer:

# tc qdisc add dev eth1 root handle 1: cbq bandwidth 10Mbit allot 1514 \
  cell 8 avpkt 1000 mpu 64
 
# tc class add dev eth1 parent 1:0 classid 1:1 cbq bandwidth 10Mbit    \
  rate 10Mbit allot 1514 cell 8 weight 1Mbit prio 8 maxburst 20        \
  avpkt 1000
Standard CBQ preamble. I never get used to the sheer amount of numbers required!

Defmap refers to TC_PRIO bits, which are defined as follows:

TC_PRIO..          Num  Corresponds to TOS
-------------------------------------------------
BESTEFFORT         0    Maximize Reliablity        
FILLER             1    Minimize Cost              
BULK               2    Maximize Throughput (0x8)  
INTERACTIVE_BULK   4                               
INTERACTIVE        6    Minimize Delay (0x10)      
CONTROL            7                               

The TC_PRIO.. number corresponds to bits, counted from the right. See the pfifo_fast section for more details how TOS bits are converted to priorities.

Now the interactive and the bulk classes:

# tc class add dev eth1 parent 1:1 classid 1:2 cbq bandwidth 10Mbit     \
  rate 1Mbit allot 1514 cell 8 weight 100Kbit prio 3 maxburst 20        \
  avpkt 1000 split 1:0 defmap c0

# tc class add dev eth1 parent 1:1 classid 1:3 cbq bandwidth 10Mbit     \
  rate 8Mbit allot 1514 cell 8 weight 800Kbit prio 7 maxburst 20        \
  avpkt 1000 split 1:0 defmap 3f

The 'split qdisc' is 1:0, which is where the choice will be made. C0 is binary for 11000000, 3F for 00111111, so these two together will match everything. The first class matches bits 7 & 6, and thus corresponds to 'interactive' and 'control' traffic. The second class matches the rest.

Node 1:0 now has a table like this:

priority	send to
0		1:3
1		1:3
2		1:3
3		1:3
4		1:3
5		1:3
6		1:2
7		1:2

For additional fun, you can also pass a 'change mask', which indicates exactly which priorities you wish to change. You only need to use this if you are running 'tc class change'. For example, to add best effort traffic to 1:2, we could run this:

# tc class change dev eth1 classid 1:2 cbq defmap 01/01

The priority map at 1:0 now looks like this:

priority	send to
0		1:2
1		1:3
2		1:3
3		1:3
4		1:3
5		1:3
6		1:2
7		1:2

FIXME: did not test 'tc class change', only looked at the source.

9.5.5. Hierarchical Token Bucket

Martin Devera (<devik>) rightly realised that CBQ is complex and does not seem optimized for many typical situations. His Hierarchical approach is well suited for setups where you have a fixed amount of bandwidth which you want to divide for different purposes, giving each purpose a guaranteed bandwidth, with the possibility of specifying how much bandwidth can be borrowed.

HTB works just like CBQ but does not resort to idle time calculations to shape. Instead, it is a classful Token Bucket Filter - hence the name. It has only a few parameters, which are well documented on his site.

As your HTB configuration gets more complex, your configuration scales well. With CBQ it is already complex even in simple cases! HTB3 (check its homepage for details on HTB versions) is now part of the official kernel sources (from 2.4.20-pre1 and 2.5.31 onwards). However, maybe you still need to get a HTB3 patched version of 'tc': HTB kernel and userspace parts must be the same major version, or 'tc' will not work with HTB.

If you already have a modern kernel, or are in a position to patch your kernel, by all means consider HTB.

9.5.5.1. Sample configuration

Functionally almost identical to the CBQ sample configuration above:

# tc qdisc add dev eth0 root handle 1: htb default 30

# tc class add dev eth0 parent 1: classid 1:1 htb rate 6mbit burst 15k

# tc class add dev eth0 parent 1:1 classid 1:10 htb rate 5mbit burst 15k
# tc class add dev eth0 parent 1:1 classid 1:20 htb rate 3mbit ceil 6mbit burst 15k
# tc class add dev eth0 parent 1:1 classid 1:30 htb rate 1kbit ceil 6mbit burst 15k

The author then recommends SFQ for beneath these classes:

# tc qdisc add dev eth0 parent 1:10 handle 10: sfq perturb 10
# tc qdisc add dev eth0 parent 1:20 handle 20: sfq perturb 10
# tc qdisc add dev eth0 parent 1:30 handle 30: sfq perturb 10

Add the filters which direct traffic to the right classes:

# U32="tc filter add dev eth0 protocol ip parent 1:0 prio 1 u32"
# $U32 match ip dport 80 0xffff flowid 1:10
# $U32 match ip sport 25 0xffff flowid 1:20
And that's it - no unsightly unexplained numbers, no undocumented parameters.

HTB certainly looks wonderful - if 10: and 20: both have their guaranteed bandwidth, and more is left to divide, they borrow in a 5:3 ratio, just as you would expect.

Unclassified traffic gets routed to 30:, which has little bandwidth of its own but can borrow everything that is left over. Because we chose SFQ internally, we get fairness thrown in for free!