Table Of Contents
Table Of Contents

Differentiated Services


In the early days of the Internet, only best effort service was defined. The Internet delivers individually each packet, and delivery time is not guaranteed, moreover packets may even be dropped due to congestion at the routers of the network. It was assumed that transport protocols, and applications can overcome these deficiencies. This worked until FTP and email was the main applications of the Internet, but the newer applications such as Internet telephony and video conferencing cannot tolerate delay jitter and loss of data.

The first attempt to add QoS capabilities to the IP routing was Integrated Services. Integrated services provide resource assurance through resource reservation for individual application flows. An application flow is identified by the source and destination addresses and ports and the protocol id. Before data packets are sent the necessary resources must be allocated along the path from the source to the destination. At the hops from the source to the destination each router must examine the packets, and decide if it belongs to a reserved application flow. This could cause a memory and processing demand in the routers. Other drawback is that the reservation must be periodically refreshed, so there is an overhead during the data transmission too.

Differentiated Services is a more scalable approach to offer a better than best-effort service. Differentiated Services do not require resource reservation setup. Instead of making per-flow reservations, Differentiated Services divides the traffic into a small number of forwarding classes. The forwarding class is directly encoded into the packet header. After packets are marked with their forwarding classes at the edge of the network, the interior nodes of the network can use this information to differentiate the treatment of packets. The forwarding classes may indicate drop priority and resource priority. For example, when a link is congested, the network will drop packets with the highest drop priority first.

In the Differentiated Service architecture, the network is partitioned into DiffServ domains. Within each domain the resources of the domain are allocated to forwarding classes, taking into account the available resources and the traffic flows. There are service level agreements (SLA) between the users and service providers, and between the domains that describe the mapping of packets to forwarding classes and the allowed traffic profile for each class. The routers at the edge of the network are responsible for marking the packets and protect the domain from misbehaving traffic sources. Nonconforming traffic may be dropped, delayed, or marked with a different forwarding class.

Implemented Standards

The implementation follows these RFCs below:

  • RFC 2474: Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers
  • RFC 2475: An Architecture for Differentiated Services
  • RFC 2597: Assured Forwarding PHB Group
  • RFC 2697: A Single Rate Three Color Marker
  • RFC 2698: A Two Rate Three Color Marker
  • RFC 3246: An Expedited Forwarding PHB (Per-Hop Behavior)
  • RFC 3290: An Informal Management Model for Diffserv Routers

Architecture of NICs

Network interface modules, such as PppInterface and EthernetInterface, may contain traffic conditioners in their input and output data path.

Network interfaces may also contain an optional external queue component. In the absence of an external queue module, Ppp and EtherMac use an internal drop-tail queue to buffer the packets while the line is busy.

Traffic Conditioners

Traffic conditioners have one input and one output gate as defined in the ITrafficConditioner interface. They can transform the incoming traffic by dropping or delaying packets. They can also set the DSCP field of the packet, or mark them other way, for differentiated handling in the queues.

Traffic conditioners perform the following actions:

  • classify the incoming packets
  • meter the traffic in each class
  • marks/drops packets depending on the result of metering
  • shape the traffic by delaying packets to conform to the desired traffic profile

INET provides classifier, meter, and marker modules, that can be composed to build a traffic conditioner as a compound module.

Output Queues

Queue components must implement the IOutputQueue module interface. In addition to having one input and one output gate, these components must implement a passive queue behaviour: they only deliver a packet when the module connected to their output explicitly requests it. (In C++ terms, the module must implement the IPassiveQueue interface. The next module requests a packet by calling the requestPacket() method of that interface.)

Simple modules

This section describes the primitive elements from which traffic conditioners and output queues can be built. The next sections shows some examples, how these queues, schedulers, droppers, classifiers, meters, markers can be combined.

The type of the components are:

  • queue: container of packets, accessed as FIFO
  • dropper: attached to one or more queue, it can limit the queue length below some threshold by selectively dropping packets
  • scheduler: decide which packet is transmitted first, when more packets are available on their inputs
  • classifier: classify the received packets according to their content (e.g. source/destination, address and port, protocol, dscp field of IP datagrams) and forward them to the corresponding output gate.
  • meter: classify the received packets according to the temporal characteristic of their traffic stream
  • marker: marks packets by setting their fields to control their further processing


When packets arrive at higher rate, than the interface can trasmit, they are getting queued.

Queue elements store packets until they can be transmitted. They have one input and one output gate. Queues may have one or more thresholds associated with them.

Received packets are enqueued and stored until the module connected to their output asks a packet by calling the requestPacket() method.

They should be able to notify the module connected to its output about the arrival of new packets.

FIFO Queue

The FifoQueue module implements a passive FIFO queue with unlimited buffer space. It can be combined with algorithmic droppers and schedulers to form an IOutputQueue compound module.

The C++ class implements the IQueueAccess and IPassiveQueue interfaces.


The other primitive queue module is DropTailQueue. Its capacity can be specified by the frameCapacity parameter. When the number of stored packet reached the capacity of the queue, further packets are dropped. Because this module contains a built-in dropping strategy, it cannot be combined with algorithmic droppers as FifoQueue can be. However its output can be connected to schedulers.

This module implements the IOutputQueue interface, so it can be used as the queue component of interface card per se.


Algorithmic droppers selectively drop received packets based on some condition. The condition can be either deterministic (e.g. to bound the queue length), or probabilistic (e.g. RED queues).

Other kind of droppers are absolute droppers; they drop each received packet. They can be used to discard excess traffic, i.e. packets whose arrival rate exceeds the allowed maximum. In INET the Sink module can be used as an absolute dropper.

The algorithmic droppers in INET are ThresholdDropper and RedDropper. These modules has multiple input and multiple output gates. Packets that arrive on gate in[i] are forwarded to gate out[i] (unless they are dropped). However the queues attached to the output gates are viewed as a whole, i.e. the queue length parameter of the dropping algorithm is the sum of the individual queue lengths. This way we can emulate shared buffers of the queues. Note, that it is also possible to connect each output to the same queue module.

Threshold Dropper

The ThresholdDropper module selectively drops packets, based on the available buffer space of the queues attached to its output. The buffer space can be specified as the count of packets, or as the size in bytes.

The module sums the buffer lengths of its outputs and if enqueuing a packet would exceed the configured capacities, then the packet will be dropped instead.

By attaching a ThresholdDropper to the input of a FIFO queue, you can compose a drop tail queue. Shared buffer space can be modeled by attaching more FIFO queues to the output.

RED Dropper

The RedDropper module implements Random Early Detection ().

It has \(n\) input and \(n\) output gates (specified by the numGates parameter). Packets that arrive at the \(i^{th}\) input gate are forwarded to the \(i^{th}\) output gate, or dropped. The output gates must be connected to simple modules implementing the IQueueAccess C++ interface (e.g. FifoQueue).

The module sums the used buffer space of the queues attached to the output gates. If it is below a minimum threshold, the packet won’t be dropped, if above a maximum threshold, it will be dropped, if it is between the minimum and maximum threshold, it will be dropped by a given probability. This probability determined by a linear function which is 0 at the minth and maxp at maxth.


The queue length can be smoothed by specifying the wq parameter. The average queue length used in the tests are computed by the formula:

\[avg = (1-wq)*avg + wq*qlen\]

The minth, maxth, and maxp parameters can be specified separately for each input gate, so this module can be used to implement different packet drop priorities.


Scheduler modules decide which queue can send a packet, when the interface is ready to transmit one. They have several input gates and one output gate.

Modules that are connected to the inputs of a scheduler must implement the IPassiveQueue C++ interface. Schedulers also implement IPassiveQueue, so they can be cascaded to other schedulers, and can be used as the output module of IOutputQueue’s.

There are several possible scheduling discipline (first come/first served, priority, weighted fair, weighted round-robin, deadline-based, rate-based). INET contains implementation of priority and weighted round-robin schedulers.

Priority Scheduler

The PriorityScheduler module implements a strict priority scheduler. Packets that arrived on in[0] has the highest priority, then packets arrived on in[1], and so on. If more packets available when one is requested, then the one with highest priority is chosen. Packets with lower priority are transmitted only when there are no packets on the inputs with higher priorities.

PriorityScheduler must be used with care, because a large volume of higher packets can starve lower priority packets. Therefore it is necessary to limit the rate of higher priority packets to a fraction of the output datarate.

PriorityScheduler can be used to implement the EF PHB.

Weighted Round Robin Scheduler

The WrrScheduler module implements a weighted round-robin scheduler. The scheduler visits the input gates in turn and selects the number of packets for transmission based on their weight.

For example if the module has three input gates, and the weights are 3, 2, and 1, then packets are transmitted in this order:

A, A, A, B, B, C, A, A, A, B, B, C, ...

where A packets arrived on in[0], B packets on in[1], and C packets on in[2]. If there are no packets in the current one when a packet is requested, then the next one is chosen that has enough tokens.

If the size of the packets are equal, then WrrScheduler divides the available bandwith according to the weights. In each case, it allocates the bandwith fairly. Each flow receives a guaranteed minimum bandwith, which is ensured even if other flows exceed their share (flow isolation). It is also efficiently uses the channel, because if some traffic is smaller than its share of bandwidth, then the rest is allocated to the other flows.

WrrScheduler can be used to implement the AFxy PHBs.


Classifier modules have one input and many output gates. They examine the received packets, and forward them to the appropriate output gate based on the content of some portion of the packet header. You can read more about classifiers in RFC 2475 and RFC 3290.

The inet.networklayer.diffserv package contains two classifiers: MultiFieldClassifier to classify the packets at the edge routers of the DiffServ domain, and BehaviorAggregateClassifier to classify the packets at the core routers.

Multi-field Classifier

The MultiFieldClassifier module can be used to identify micro-flows in the incoming traffic. The flow is identified by the source and destination addresses, the protocol id, and the source and destination ports of the IP packet.

The classifier can be configured by specifying a list of filters. Each filter can specify a source/destination address mask, protocol, source/destination port range, and bits of TypeOfService/TrafficClass field to be matched. They also specify the index of the output gate matching packet should be forwarded to. The first matching filter determines the output gate, if there are no matching filters, then defaultOut is chosen.

The configuration of the module is given as an XML document. The document element must contain a list of <filter> elements. The filter element has a mandatory @gate attribute that gives the index of the gate for packets matching the filter. Other attributes are optional and specify the condition of matching:

  • @srcAddress, @srcPrefixLength: to match the source address of the IP
  • @destAddress, @destPrefixLength:
  • @protocol: matches the protocol field of the IP packet. Its value can be a name (e.g. “udp”, “tcp”), or the numeric code of the protocol.
  • @tos,@tosMask: matches bits of the TypeOfService/TrafficClass field of the IP packet.
  • @srcPort: matches the source port of the TCP or UDP packet.
  • @srcPortMin, @srcPortMax: matches a range of source ports.
  • @destPort: matches the destination port of the TCP or UDP packet.
  • @destPortMin, @destPortMax: matches a range of destination ports.

The following example configuration specifies

  • to transmit packets received from the 192.168.1.x subnet on gate 0,
  • to transmit packets addressed to port 5060 on gate 1,
  • to transmit packets having CS7 in their DSCP field on gate 2,
  • to transmit other packets on defaultGate.
  <filter srcAddress="" srcPrefixLength="24" gate="0"/>
  <filter protocol="udp" destPort="5060" gate="1"/>
  <filter tos="0b00111000" tosMask="0x3f" gate="2"/>

Behavior Aggregate Classifier

The BehaviorAggregateClassifier module can be used to read the DSCP field from the IP datagram, and direct the packet to the corresponding output gate. The DSCP value is the lower six bits of the TypeOfService/TrafficClass field. Core routers usually use this classifier to guide the packet to the appropriate queue.

DSCP values are enumerated in the dscps parameter. The first value is for gate out[0], the second for out[1], so on. If the received packet has a DSCP value not enumerated in the dscps parameter, it will be forwarded to the defaultOut gate.


Meters classify the packets based on the temporal characteristics of their arrival. The arrival rate of packets is compared to an allowed traffic profile, and packets are decided to be green (in-profile) or red (out-of-profile). Some meters apply more than two conformance level, e.g. in three color meters the partially conforming packets are classified as yellow.

The allowed traffic profile is usually specified by a token bucket. In this model, a bucket is filled in with tokens with a specified rate, until it reaches its maximum capacity. When a packet arrives, the bucket is examined. If it contains at least as many tokens as the length of the packet, then that tokens are removed, and the packet marked as conforming to the traffic profile. If the bucket contains less tokens than needed, it left unchanged, but the packet marked as non-conforming.

Meters has two modes: color-blind and color-aware. In color-blind mode, the color assigned by a previous meter does not affect the classification of the packet in subsequent meters. In color-aware mode, the color of the packet can not be changed to a less conforming color: if a packet is classified as non-conforming by a meter, it also handled as non-conforming in later meters in the data path.


Meters take into account the length of the IP packet only, L2 headers are omitted from the length calculation. If they receive a packet which is not an IP datagram and does not encapsulate an IP datagram, an error occurs.


The TokenBucketMeter module implements a simple token bucket meter. The module has two output, one for green packets, and one for red packets. When a packet arrives, the gained tokens are added to the bucket, and the number of tokens equal to the size of the packet are subtracted.

Packets are classified according to two parameters, Committed Information Rate (\(cir\)), Committed Burst Size (\(cbs\)), to be either green, or red.

Green traffic is guaranteed to be under \(cir*(t_1-t_0)+8*cbs\) in every \([t_0,t_1]\) interval.


The SingleRateThreeColorMeter module implements a Single Rate Three Color Meter (RFC 2697). The module has three output for green, yellow, and red packets.

Packets are classified according to three parameters, Committed Information Rate (\(cir\)), Committed Burst Size (\(cbs\)), and Excess Burst Size (\(ebs\)), to be either green, yellow or red. The green traffic is guaranteed to be under \(cir*(t_1-t_0)+8*cbs\), while the green+yellow traffic to be under \(cir*(t_1-t_0)+8*(cbs+ebs)\) in every \([t_0,t_1]\) interval.


The TwoRateThreeColorMeter module implements a Two Rate Three Color Meter (RFC 2698). The module has three output gates for the green, yellow, and red packets.

It classifies the packets based on two rates, Peak Information Rate (\(pir\)) and Committed Information Rate (\(cir\)), and their associated burst sizes (\(pbs\) and \(cbs\)) to be either green, yellow or red. The green traffic is under \(pir*(t_1-t_0)+8*pbs\) and \(cir*(t_1-t_0)+8*cbs\), the yellow traffic is under \(pir*(t_1-t_0)+8*pbs\) in every \([t_0,t_1]\) interval.


DSCP markers sets the codepoint of the crossing packets. The codepoint determines the further processing of the packet in the router or in the core of the DiffServ domain.

The DscpMarker module sets the DSCP field (lower six bit of TypeOfService/TrafficClass) of IP datagrams to the value specified by the dscps parameter. The dscps parameter is a space separated list of codepoints. You can specify a different value for each input gate; packets arrived at the \(i^{th}\) input gate are marked with the \(i^{th}\) value. If there are fewer values, than gates, then the last one is used for extra gates.

The DSCP values are enumerated in the DSCP.msg file. You can use both names and integer values in the dscps parameter.

For example the following lines are equivalent:

**.dscps = "EF 0x0a 0b00001000"
**.dscps = "46 AF11 8"

Compound modules


The AFxyQueue module is an example queue, that implements one class of the Assured Forwarding PHB group (RFC 2597).

Packets with the same AFx class, but different drop priorities arrive at the afx1In, afx2In, and afx3In gates. The received packets are stored in the same queue. Before the packet is enqueued, a RED dropping algorithm may decide to selectively drop them, based on the average length of the queue and the RED parameters of the drop priority of the packet.

The afxyMinth, afxyMaxth, and afxyMaxp parameters must have values that ensure that packets with lower drop priorities are dropped with lower or equal probability than packets with higher drop priorities.


The DiffservQueue is an example queue, that can be used in interfaces of DS core and edge nodes to support the AFxy (RFC 2597) and EF (RFC 3246) PHB’s.


The incoming packets are first classified according to their DSCP field. DSCP’s other than AFxy and EF are handled as BE (best effort).

EF packets are stored in a dedicated queue, and served first when a packet is requested. Because they can preempt the other queues, the rate of the EF packets should be limited to a fraction of the bandwith of the link. This is achieved by metering the EF traffic with a token bucket meter and dropping packets that does not conform to the traffic profile.

There are other queues for AFx classes and BE. The AFx queues use RED to implement 3 different drop priorities within the class. BE packets are stored in a drop tail queue. Packets from AFxy and BE queues are sheduled by a WRR scheduler, which ensures that the remaining bandwith is allocated among the classes according to the specified weights.