There is quite some confusion about QoS terms, and different sources give different interpretations to them. While the terms used here are also not ‘the reference’ they are at least supposed to be consistently used within the scope of this site.

PDU Protocol Data Units

Throughout this site frames and packets are used interchangeably. They designate an arbitrarily sized ‘Protocol Data Unit’ which is the official term for the data transmission units on an unspecified protocol.

Officially frames designate variably sized PDUs in Ethernet, Frame Relay and MPLS networks.

Packets are typically used in IP context.

Traffic Conditioning Actions

Traffic conditioning actions (TCA) are the superset of regulatory policies enacted on traffic flows in order to enforce an SLA. This includes policing, shaping and marking.

Policing is a means to enforce a time-based upper bound on what a data source can inject into the network. It is purely based on metering, and discarding or marking packets based on the metering result. It does not involve queuing and does not require storage. Policing does not introduce delay. A policer either discards a packet or lets it pass without altering its timing.

Shaping is a means to enforce a time-based upper bound bound on what a data source can inject into the network. When a newly arriving packet is above the upper bound defined by the shaping parameters packets will be placed into the shaped queue and delayed. Packets are only discarded when the shaped queue overflows. Shaping alters the arrival curve of traffic flows. Shaping introduces delay.

Marking is a means to convey subsequent nodes in the connectionless network the treatment a PDU deserves. It can indicate both traffic priority and discard eligibility within a traffic class. Marking can be a result of policing, or simply marking based on the source or based on other criteria such as destination, protocol information or a combination of multiple conditions.

Burst

A traffic stream bursts when it sends data faster than its declared or allowed limit. The comparison can be done to its long term average, but typically the relevant comparison is with its Service Level Agreement. As such, as traffic streams bursts when it sends traffic faster than the CIR of the policer that is used to limit it, or the shaping rate of the shaper that limits it.

Bursts can have multiple definitions. draft-ietf-bmwg-lanswitch-01 defines packets belonging to a burst if they arrive back to back on the Ethernet physical interface. An alternative approach is to consider a set of packets in the traffic stream. If the amount of data transmitted by the packets divided by the time delta between the complete reception of the first and the last packet of the packet train is above the CIR of the stream, then the packet train is a burst. This alternative definition is used on this site.

Another possible option to define the length of a burst is between the first packet that causes CBS of the policer to decline until the last packet from which the token count consistently (monotonously) grows back to a level that exceed the starting level of the token count when the burst started.

 Either way, in most practical cases it is inappropriate to claim that traffic bursts at X or Y rate because depending on the time frame that we look at, the burst has a different ‘rate’ and would cause a different backlog in the shaper, or a different decline in the token count of a policer. In general bursts are considered because they would cause overflow in a shaper or underflow in a token bucket policer.