This document discusses rate control for multimedia streaming in high bandwidth environments. It covers feedback information used in rate control such as packet delivery and loss statistics. Non-linear theory is discussed where the total rate is made up of a constant source rate and a feedback rate used to control the source rate based on packet dropping. Recent developments and challenges are presented such as delayed feedback problems. Solutions proposed include using back-to-back packets for startup rate detection and predicting network conditions through techniques like ARMAX and neural networks or analyzing group traffic behavior.

1. Rate Control for
Multimedia Streaming in
High Bandwidth Environment - 3
Fitri Setyorini
Nakazato Lab

2. Content
 Rate control definition
 Non Linear Theory
 Recent Development of research
 Solution proposal

3. Facts about Rate Control
 End to end protocol
 It is used to deliver UDP packet, via RTP as the
carriage
 in certain speed based on the information
 The information used is RTCP packet, which is
feedbacked through the source

4. Rate Control
SOURCE Receiver

5. So..we need to keep the traffic

6. Not to be like this

7. This is what
RATE CONTROL
Do

8. How to control it ?
 We have information from the feedback,
− What kind of information available ?
 We have to use equation to control the speed,
− What kind of equation we will use ?

9. Feedback Information
 1. Number of packet delivered at receiver
 2. Loss occurred at receiver
 3. Time needed to travel from source to receiver
(RTT)

10. Non Linear Theory
Total rate defines the rate from the source
Constitute of the
− constant source rate, Ini(k) ;
− feedback rate, U(k).
We use feedback to control rate
If packet dropping occurs, the feedback rate is negative
and source will reduce its rate.
If no packet dropping, the feedback rate is positive and
the source will increase the rate
We called it non linear because the theory predict
the packet accumulation non linearly

11. Dumbbell topology

12. Objectives
 No Drop
 Higher Throughput

13. State of Research
 We still use priority for RTCP packet on the router
− Without priority, RTCP packet will be dropped just
like RTP packet, therefore the rate control won't work
efficiently
− New time out mechanism is necessary to smooth out
some lost RTCP packets
 Start up mechanism had not been decided
− Before RTCP packet is received, we can not control the
speed
− Any idea ?

14.  Delayed feedback problem
− If the propagation delay is too big, then the
information will be too late to be received and
processed. The feedback will be obsolete, because by
that time the network condition already change
− Can not be avoided because this is the nature of
network

15. Drop : buffer 200,delay 10ms

16. Drop : buffer 200,delay 200ms

17. Throughput : buffer 200,delay 10ms

18. Throughput : buffer 200,delay 200ms

19. Solution to Delayed feedback problem

Simplest thing to do :
Increase RTCP packets ahead, to avoid lack of
information which cause losses

20. 1 RTCP/RTT

21. 2 RTCP/RTT

22. 4 RTCP/RTT

23. Shortcomings :
 If we sent RTCP after drop happens, it will be too
late to prevent dropping
 We can schedule some RTCP packet to be sent
ahead in the prescribed interval by looking at
RTT, but the questions are
− How many RTCP is necessary ?
− When we need to launch RTCP or what is the ideal
RTCP interval ?
 We do not want to flood the network with RTCP
packet

24. Solution (1/2)
 Start Up problem :
− We use back to back packet to detect the bottleneck
link
− We have bottleneck bw, we can start with it ->
SOLVED
 Back to back (B2B) packet : 1500 bytes
 We measure interval between 2 consecutive B2B
 Bottleneck bandwidth = (1500*8)/interval
 As initial state, we use B2B bandwidth as RTP
rate

25. Solution (2/2)
 Delayed Feedback problem :
− The RTCP maybe too late whilst network condition
change.
− Why don't we predict the network condition to prevent
dropping
− Prediction by the following method
 Combining ARMAX algorithm and neural network or
 Analyzing group behaviour of group flow

26. Combining ARMAX with
Neural Network
 The author of non linear already presume the
problem with variable delay, therefore he propose
a more complicated NN to deal with delay
 NN can have unpredictable result (which is why
we do not prefer)
 NN is used to find the parameter for accumulated
traffic
 What kind of NN suitable for our model ?

27. Analysing Group Behaviour
 Dropping occurs because all flows put high
number of packet in the network
 Dropping in one flow reflects all other flow
condition
 Using this information, we can predict group
behavior
 Group behavior will be used to detect when we
should launch RTCP packet

28. Thank you for your attention