From: Ethan Phelps-Goodman (ethanpg@cs.washington.edu)
Date: Tue Oct 26 2004 - 18:05:35 PDT
Real-Time Apps in ISPNs
Clark, Shenker and Zhang
This paper presents an framework for reasoning about real time systems, and
a proposed queuing system for QoS guarantees. The paper is probably most
useful for how it frames the discussion. Service guarantees, they point out,
had been too inflexible in past literature. Most real-time applications,
such as voice and video, don't need an absolute bound on delay, but instead
want the lowest possible average delay. They can get by without a hard upper
bound because of buffering at the receiving end, and adaptability of the
encoding.
After a good discussion of the nature of service guarantees, they describe
the mechanisms they have built for achieving these guarantees. They start
with WFQ, which was invented elsewhere. They show that WFQ along with a
token bucket specification of service requirements is good for providing
isolation, but that this isn't necessarily what is wanted. In particular, it
is reasonable to temporarily give a bursty source more of its fair bandwidth
in order to reduce the average jitter. They also discuss the cumulative
effects of jitter from multiple switches, and suggest adding an explicit
jitter count to the packet header. This way, packets which have been delayed
longer than average can jump to the head of the queue in later switches.
Overall, the touch on most, if not all, the relevant issues. One piece
that's missing, which the discuss in the conclusion, is that without a
payment system built in, there is no reason for a source to ever request
anything less than the highest level of service.
Ethan
This archive was generated by hypermail 2.1.6 : Tue Oct 26 2004 - 18:05:40 PDT