From: Seth Cooper (scooper@cs.washington.edu)
Date: Wed Oct 27 2004 - 01:19:32 PDT
This paper discusses the framework needed to support an Integrated
Services Packet Network (ISPN) on the Internet. This network must
support real time applications. The paper analyzes real time
applications as playback applications, where one side encodes and sends
data to the other side, which must decode the data before a certain
playback point. It then observes that there are two basic types of real
time applications: rigid, in which the playback point is always fixed,
and adaptive, where the playback point can be adjusted in accordance
with network delay. Each type of application requires a different type
of service from the network: rigid requires "guaranteed" service from
the network, while adaptive requires "predictive". The paper then
discusses the building blocks necessary to support these types of
service as well as the standard datagram protocol.
One interesting part of this paper is the discussion of WFQ vs FIFO,
and how WFQ achieves isolation, but FIFO achieves sharing. Although WFQ
is required for guaranteed service, it is bad for predictive service.
This is because a FIFO scheduling algorithm will share the delay from
multiple sources, thus reducing the jitter for all of them. It would
also be possible to add priority classes to predictive service. Higher
priorities would have lower jitter; lower priorities would absorb the
jitter of the higher priorities, but the service would come at a reduced
cost.
A weakness of the paper are the criteria presented for admission
control. The paper merely states that when reserving resources, only
10% needs to be reserved for datagram traffic. The claim is stated to
be ad hoc, and no facts are presented to back it up. 10% seems awfully
low; imagine if the Internet's bandwidth were cut to 10% of what it is
currently. If there is still a large number of datagram traffic, it may
be difficult to decide exactly how many resources need to be reserved
for it.
This paper remains relevant because it provides an architecture for
Quality of Service in the Internet. It also presents an effective
scheme for allowing users to pay different costs for different levels of
service, as priority classes in predictive service.
This archive was generated by hypermail 2.1.6 : Wed Oct 27 2004 - 01:19:33 PDT