From: Danny Wyatt (danny@cs.washington.edu)
Date: Tue Oct 26 2004 - 21:55:33 PDT
Supporting Real-Time Applications in an Integrated Services Packet
Network: Architecture and Mechanism
David D. Clark, Scott Shenker, Lixia Zhang
This paper presents a method for commingling service commitments for
real-time traffic as well as non-real-time traffic in a packet switched
network. Service commitments are defined by rate and delay, and they
divide these commitments into two classes: guaranteed and predicted.
Guaranteed commitments are for rigid applications intolerant of any
variation in delay beyond a given bound. Predicted commitments are for
applications that only need mostly-real-time service and can tolerate
unbounded fluctuations around a mean delay. They reiterate that
weighted fair queueing can provide guaranteed commitments, but maintain
that it is too rigid for predicted commitments. For predicted
commitments, good old FIFO queuing turns out to be the best---but only
across one link.
For multiple links, they devise a new scheme---FIFO+---that maintains a
packets expected delay and schedules over-delayed packets ahead of
others. This seems to be a crucial point in their architecture, and I
think they give it short shrift (especially considering how prolix they
are overall). Which brings me to my second criticism: they advertise
and "architecture and mechanism" but aside from the 3 paragraph appendix
describing a simple simulation, there is no more thorough "mechanism" to
be found. Overall, the paper suffered from an excess of ambition and a
shortage of detail.
That said, I appreciated the insight that isolation and sharing are
solutions to two related but not identical problems, and I admire the
attempt to combine multiple classes of reliability into a single
scheduling algorithm.
This archive was generated by hypermail 2.1.6 : Tue Oct 26 2004 - 21:56:30 PDT