Modeling Resource Sharing Dynamic of VoIP usersOver a WLAN Using a Game-Theoretic Approach

Presented by Jaebok Kim

Motivation

A great number of people using Internet for multimedia contents today

Need resource sharing mechanism & congestion control for limited resource and competition

Suggested many models & mechanismfor success of multimedia contents services in Internet However, suggested models will not be applied widely as TCP was

Why?

Motivation

450 M users

Because of complexity of Internet in terms of a number of applications, competing developers, users

So, we can dare to say….

Motivation

Network doesn’t implement resource sharing mechanism!

Instead, where can we find this?

Yes, end-user applications do it.called, end-user congestion controlOnly widely deployed multimedia congestion control

Motivation

Yes, it’s Skype, again.

So, resource sharing mechanism is delegated to end-users.

What can they do for their quality?

What if they choose their strategy freely?

Motivation

Is it bad or good for a network to allow users to havethat freedom?

What are the performance consequences of letting the users to freely choose their rates? (how good? how bad?)

Rising Questions

How to get the answers

Conducting experimental measurement to see per-formance consequence

Conflicts occur because of limited resource between users!There is no technical solution!Adopting this view, no technical solution!

We’ll have real experiment letting users have freedom.It will show collapse or adaptation of users.

That’s it? is it done?

No, we need to explain the result.

So, we’ll use a modelWhy?

Why do we use model?To apply quantitative reasoning to observations To discover aspect of reality or understand it

How do we usually start?

characterizing the system assumptions about how it worksinterpret these into equations & a simulationvalidation:

So, to understand what if users have freedom to choose

strategy, we need a model.

What kind of model do we need?

Its goal is to understand interaction between players in competi-tion.

Let’s adopt Nash Equilibrium, one of its concepts.

No player can benefit by changing strategy while the other players keep theirs unchanged

It’s similar to that users converge eventually.

Two types of Game Theory

A normative theory – to find what designers should do to solve a problem

A descriptive theory – to understand reality and infer metrics(called behavioral game theory)Give us better understanding of the behaviors of users, competing for medium access in a WLAN.

How different from previous work

Plus, wireless link

Active measurement(emulating the real measurement) using Traffic Generator & NS

Behavioral Model (MC)

Transition Rate be-tween two states based on measure-ment of voice quality

State: the vector of n of players adopting strate-gies(#users low rate,# users average rate, #users high rate)Ex) (1,2,3) , (0,6,0)…

Choosing high rate doesn’t mean high quality of voice!Ex) (0,0,6)

The example of MC

Let’s see the simple example.Only two users, two choices of rate (high, low)We’ll make a model looking like this.

(2, 0) (0, 2)(1, 1)0.4 0.6

0.30.1

0.70.6 0.3

Injecting data,Computing statistics

Giving flexibility for sce-narios

Estimating MOS(mean opinion score)

:1~5

Using PSQA(Pseudo-Subjective Quality

Assessment)

Transition Rate calculated from suggested MOS

RNNlearning algorithm

,better than manual mea-surement,

using six parameters(loss rate…)

Emulating the behavioral measurementwith exact flow mapped to VivaVoz’s jitter, loss..

Overview of the model

Emulating the behavioral measurement

We need data to parameterize the model.

Sum of user’s rate

Swings around 160 kbps

Emulating the behavioral measurementusing Traffic Generator

Channel Performance by Active Mea-surement

Now, we can parameterize the QoS model by using the data.

Sim & measurement are matched well. So,We can use A.M to parameterize our model.

Active measurement(emulating the real measurement) using Traffic Generator & NS

Behavioral Model (MC)

State: the vector of n of players adopting strategies(#users low rate,# users average rate, #users high rate)changing to another depending on the quality of service

So far, we’ve got parameters for QoS model.What about parameters for MC model?

The Top Layer: MC model

N of users adopting L * PDF

Makes a mistakeWorse than before

Try different rate

After change, quality becomes better than before orBetter than toler-ance level

A vector of a measure of voice quality A measure voice quality, when any user adopts strategy lA measure of voice quality, after changing strategy from l to m

Transition Rates between states in MC

T & ɣ are extra parameters suggested in the paper.We assume users can’t tolerate a certain poor quality.T: exploring rate ɣ: tolerance levelWe know most parameters calculated from QoS model,

But, how do we decide ɣ?Let’s wait until we take a look around the real experiment.

Example of transition in MC

(2, 0) (0, 2)(1, 1)? ?

??

?? ?

S(#High, #Low), initial state is (2,0)How do we get the transition rate for this,when a user change strategy from high rate to low rate?S(2,0) ->S(1,1)

First, we get the a vector of MOS for them from QoS model.Let’s assume that U(2,0) & U(3,3) for each state.U2 after the change is 3, it’s greater than U1, 2 before the change. So, the transition rate is calculated from the following:S1 * = 2 * 0.4(assumed value) = 0.8

0.8

So far, we’ve understood the structure of the game theory model &

how it works.

But after we see the result of real experiment, we’ll see the result

pointed out by the model.

By comparing them, we know whether the model is correct or not.

So far, what have we done?

Free to choose rate, no bail out!

Non Interactive: Pre-recorded voiceInteractive: Freely talked

Share AP, Choose rate freely

3 pairs of users

Look at the final rates, selected at the end of experiment.We could see “auto-regulation ability”.

Without any technical solution, they converged to an efficient state(Using resources almost fully!) No collapse!

Tend to be more aggressive when using interactive application

Aggregate rate varies from 130 to 160kbpsIn this range, MOS is greater than at least 2.6

How to set tolerance level: r ?

Yes, the results of experiments and the model are so close.The model shows the adaptation of users.

N of All possible states = 28Since 6 hosts & 3 rates

short term:first transition

medium term:frequently changing rates

long term: reach N.E for the first time

Super Long term:steady state

For simplicity tolerance level is set to zero.A way of initial state(0,5,1) to 28 possible states

Without setting the tolerance level, the possibility of going to tragedy of the commons will be high.

P((0,5,1)->(0,6,0))

End-user dynamic resource sharing mechanism is a big issue.

We ask a question what happens if users are allowed to have freedom.

We could see the result by conducting experimental measure-ment.

We expanded game-theory model to explain the adaptation of users.

We compared the experiment with the model.

“Users could adapt to the network conditions and converged to efficient equilibrium.”

1) Extending the previously suggested model for behavioral ex-periments and getting the method to parameterize it

2) Conducting the behavioral experiments and showing how to use the model in order to explain the results of the experi-ment.

Contribution of this paper

Thank you

Q&A

Users tried to utilize rate fullyand later, began adaptation to share resources

With tolerance level 2.6

Tolerance level is set to 2.6The probability of all possible initial states to specific final states.

Start with high rates is worse to get the equilibrium.

Parameters

How to train RNN & use it?

Using VivaVoz(conducting real tests) & NetEM(emulating the network conditions like loss rate)

People count up to 20 as quickly as possible.

Free talking about a picture and scenario

Individual gives quality score.

The results as MOS form database for RNN to train.

After training, RNN can produce MOS results by using parameters automatically.

Much more cost-effective than real manual measurement!

Training of RNN

Output of training