From: Susumu Harada (harada@cs.washington.edu)
Date: Mon Oct 25 2004 - 13:34:12 PDT
"Supporting Real-Time Applications in an Integrated Services Packet
Network: Architecture and Mechanisms"
D. Clark, S. Shenker, and L. Zhang
The paper presents a network architecture called the Integrated Services
Packet Network (ISPN) which provides a mechanism for supporting both
guaranteed and predicted services for real-time applications (as well as
best effort services), with focus on play-back applications where a source
host streams some media to be played back at the destination host where
buffering and play-back point adjustment are used in an attempt to
recreate the stream at the original rate despite jitters introduced by
network delays.
The key idea of the paper is that it combines several scheduling
algorithms into one scheme that allows the network to support all three
types of services mentioned above. It does this by using Weighted Fair
Queuing (WFQ) as the overarching scheduling algorithm, which provides
isolation among the guaranteed service flows plus one more flow, which is
the flow containing all the predicted and best effort service flows.
The predicted and best effort service flows are scheduled within that
particular WFQ flow by the use of FIFO+ with priority classes, in which
the best effort service flow is assigned the lowest priority. The idea
behind FIFO+ algorithm is that packets that are running really late due to
accumulated delays along a multi-hop path are given higher priority and
thus sent first at a particular router even if it arrives later than some
other packet with lower accumulated delay.
I thought that the idea of embedding one scheduling policy within another
to support multiple real time application needs was interesting, but one
concern I have is the scalability of this solution as the number of flows
increase and in situations when the ratio of the usage of the network by
real-time applications versus non-real-time applications fluctuate a lot.
I would like to know in such situations, whether the router will be able
to handle the processing of all the flows and what the resulting network
performance will be. There also seems to be a problem that in order for
this architecture to provide real value, all the routers in the network
will have to be upgraded with this capability, a very unlikely scenario.
This archive was generated by hypermail 2.1.6 : Mon Oct 25 2004 - 13:34:13 PDT