1. Is Dynamic Multi-Rate Worth the Effort?
Matthew Peace
Wireless Information Networking Group

2. Problem
With the progression toward continuous media distribution, end systems are expected to be cooperative in determining/adapting transmission rates
In a multi-rate multicast, several desirable fairness properties can be achieved

3. Problem
A static multi-rate system may only be able to support coarse grained adaptation
Using a dynamic multi-rate system might be more reasonable

4. Purpose of Paper
To find the possible benefit of a dynamic multi-rate multicast solution with respect to inter-receiver fairness
Does this gain compensate for higher implementation costs associated with a coding scheme and dynamic partitioning?

5. Inter-Receiver Fairness
“satisfaction” of a receiver given by ui utility function, which is a function of gj (its actual rate) and theoretical fair allocation ri
Optimality occurs when
gj = ri

6. Receiver Utility Function

7. Inter-Receiver Fairness
Goal in this system is to maximize collective satisfaction of receivers of a multicast session (sum of the utility values)
Uj = Utility for set of receivers Gj is maximized when the worst receiver’s
gj = ri

8. Inter-Receiver Fairness
Where lj is the the data rate of the jth layer
L = number of layers (groups) in a session
N = number of receivers in a session
Ri = the ith receiver
ri = the theoretical fair allocation for Ri
Gj = the set of receivers subscribed to layers 1 to j
nj = the number of receivers in Gj
gj = the cumulative data rate in Gj

9. Inter-Receiver Fairness
US = the session Utility (sum over layers)

10. Protocol Issues Optimal Rate Estimation
Receivers determine when to send feedback
Optimal rate derived from an equation modeling long term TCP throughput
Measure the Lost Rate Event
Measure the Round Trip Time

11. Protocol Issues (Cont.)
Feedback Suppression
Receiver is allowed to send feedback only if it’s utility degradation (uopt – ui) exceeds a certain threshold
Δ ui = uopt X (1 – αj)
αj derivation
Size of Gj must be taken into consideration
The larger the size of Gj the higher the utility degradation must be in order to send feedback
αj is a function of the number of receivers in G

12. Protocol Issues (Cont.)
Avoiding Leave Action
If a receiver calculates the theoretical rate to be less than the current receiving rate, it may leave the highest layer immediately
To avoid coarse-grained quality degradation:
Over a time interval T, the sender is collecting receiver feedback for each layer
Each receiver calculates ri . If ri < gj a report is sent to the sender and the receiver waits for the next announcement of the sending rates
Only if new rate has not been lowered to accommodate a receiver’s reported rate, then the receiver is forced to leave a group

13. Experiments Generation of 500 rates, with min and max rates set
The inter-receiver fairness is maximized when all receivers are served optimally or

14. Experiments (Cont.)
“Goodness” of a session determined by ratio of US to USopt

15. Results Single-Rate vs. Multi-rate
Optimality occurs when number of layers L approaches number of receivers N, but higher number of layers causes more overhead
First experiment studies the effect of increases the number of layers for a logarithmic utility function and a linear utility function

16. Results (Cont.) Normal Distribution of receiver rates mean = 1248kbps
and varying standard deviation = 2k

17. Results (Cont.)
Uniform Distribution with varying range [rmin, 2k X rmin]

18. Results (Cont.) Static Layers vs. Dynamic Layers
Rmin and Rmax used to calculate predefined static layer rates and compared with situation where adaptive techniques are used on the actual distribution of the rates

19. Results (Cont.)

20. Results (Cont.)
Assumed a trimodal distribution to find the rates for the static layers.
Then, the effect of receivers drifting from last mode to an additional one was simulated

21. Results (Cont.)

22. Conclusions
Adaptation in a multi-rate multicast mode may increase the overall satisfaction with only a few layers in an unpredictable environment, thus offsetting the cost of added complexity to the system