next up previous contents
Next: Detailed Analysis of the Up: Network Service Models Previous: RSVP

Service Classes and Assurance

As explained in chapter one, there are a variety of proposed service classes in the Integrated Services Internet, ranging from Best Effort, through Controlled Load, to Guaranteed Service.

Associated with these are two functions:

1.
Admission Control Decision - whether traffic can be supported with the current network resources.
2.
The Policing Action - monitoring of a flow may determine that it exceeds either what has been asked for or what can be supported, or both.

Admission control can also be viewed as ``refusal control'' or call or reservation blocking. Thus for Best Effort, there is no admission test, and policing consists simply of packet dropping. There is much debate right now about exactly what packet dropping policy is fair to a set of competing best effort flows, under overload conditions. it is the case that FIFO (or ``Drop Tail'') is not fair to flows which adapt based on monitored successful rates of packets per RTT (e.g. TCP), with different RTTs. Other schemes such as Random Early Detection routing[#!red!#], promise better fairness, and even help discern deliberate attempts to gain an unfair share.

For services with guaranteed performance, in general, we would expect a network to be designed to admit more calls than it denies, even when there are a borad range of services available. Most users will use the average service, after all. However, the average service may be something that changes over time. It may also be the case that the user has no means to tell (at least easily) exactly what parameters should be set to what values. In fact, measurements of actual parameters show that they vary rather more, and in more long term correlated way than had previously been expected.

A consequence of this has been that people have proposed ``Measurement Based Admission Control''. This is basically an extension of the Internet philosophy of sending a packet, and ``just seeing'' if there is enough capacity.

However, we still need some way for the sender to indicate that some level of guarantee is needed. It has been suggested that this indication could take a very simple form: simply a price, plus a delay sensitivity. This is still a research topic.


next up previous contents
Next: Detailed Analysis of the Up: Network Service Models Previous: RSVP
Jon CROWCROFT
1998-12-03