From: Andrew Putnam (aputnam@cs.washington.edu)
Date: Wed Oct 27 2004 - 03:18:48 PDT
Supporting Real-Time Applications in an Integrated Services Packet
Network: Architecture and Mechanism
David D. Clark, Scott Shenker, Lixia Zhang
Summary: This paper introduces a network architecture supporting both
hard and soft real-time network traffic as well as standard datagram
traffic.
The key observation in this paper is that many real-time
applications do not require hard, worst-case guarantees, or are
tolerant of temporary disruptions. These applications can adapt to the
network conditions in order to catch up when packets are getting behind
and to throttle back when traffic gets behind. These adaptable
real-time applications flow at a lower priority than hard real-time
applications. Since these applications can adapt, they can use the
measured delay of the network rather than the worst case delay. This
allows more applications to run together in their average-case rather
than their worst case states, This increases gateway utilization and
application performance.
Another key observation in this paper is that real-time applications
need isolation of network traffic, but this is a separate issue from
sharing network traffic. Isolation of network traffic is required for
gateways to provide the timing guarantees for the separate
applications.
One major flaw with this paper is that it is extremely difficult to
test this architecture. The authors recognize this shortcoming, yet do
not provide much in a way to revise the tests to be more meaningful. A
major part of this problem is that it is extremely difficult to
determine "typical" network traffic. Particularly for 1992, there were
no good ways of testing new algorithms on an Internet-like testbed. So,
while the results presented are nice and informative, the absolute
results may be significantly different in the real world.
The authors slip in an interesting note that their mechanism will
only work if there is incentive for users to use as low a guarantee as
possible. In order to do this, there must be a good accounting system
in order to track who is using which guarantee level of traffic.
Accounting is a major modern problem that will have to be solved to
make this a viable architecture.
One of the authors' fundamental assumptions is that the real-time
applications can buffer packets that are not yet needed. This is
certainly valid for streaming applications, but it is a shaky
assumption for general real-time applications. For instance, air
traffic control systems need data as quickly as possible, but there is
not a sense of buffering until a particular play point. Instead,
packets that are young enough are used, packets that are too old are
immediately discarded.
This archive was generated by hypermail 2.1.6 : Wed Oct 27 2004 - 03:18:56 PDT