Chaos with QoS
Position Statement
Dagstuhl, June 1995
Wolfgang Effelsberg
University of Mannheim
What is Quality of Service?
Application QoS parameters
image quality, image size, frame rate, startup delay, reliability
...
Network QoS parameters
bandwidth, delay, jitter, loss rate ...
The user is unable to specify network QoS parameters.
Why is QoS a Problem?
- The relationship between application QoS parameters and network
QoS parameters is very complex:
- QoS must be end-to-end ("photon-in / photon-out"),
not TSAP to TSAP
- The best operating point might change during connections.
Providing QoS Guarantees
Service quality can be classified into
- best - effort (no QoS)
That is not what we want.
- with stochastic guarantees (e.g. for 99 % of all packets the
delay
will be < 5 ms)
We don't know how to implement that.
- with deterministic guarantees
Currently the only feasible approach.
Two application/network interface paradigms
- the contract approach
- the media scaling approach, nowadays called "adaptive
applications"
QoS Adaptation
1. The application adapts to changes in the network:
2. The network adapts to changing application requirements:
- multicast filtering
- dynamic join and leave for multicast connections
- dynamic channel management
3. both!
Media Scaling
Proposed by IBM ENC, Heidelberg and others
If two applications want to scale up, who wins?
Multicast Filtering
- Tree nodes contain filters to optimize bandwidth utilization.
- Requires hierarchical stream encoding.
- Will require application-specific knowledge in the network nodes!
RSVP proposes such filters in IP nodes. The filters themselves
are application-independent, they can be activated based on an
identifier in the packet stream.
Dynamic Join and Leave for Multicast
It is difficult to maintain QoS when membership changes in a multicast
connection.
Dynamic Channel Management
Proposed by the Tenet group at ICSI, Berkeley.
Idea: Assign an alternative channel through the network when QoS
parameters change.
Procedure: Compute alternate channel, reserve resources on that
channel, switch over in an atomic transition. New channel can
partially overlap old channel.
Can be network-initiated or application-initiated.
Conclusions
1. We need an overall QoS architecture including
- the interface
- the network components
- the end systems
(not just a list of parameters)
2. We don't know how to implement stochastic guarantees.
3. QoS mapping is difficult, must be solved for each application
separately.
4. For the time being, the contract approach with interval
parameters
[lower bound, upper bound], connection-oriented multicast
and resource reservation seems to be a feasible solution for packet-switched
internetworks.
Stephan Fischer <fisch@pi4.informatik.uni-mannheim.de>
Last modified: Tue Jun 20 02:28:24 1995