Department of Information Engineering University of Padova, ITALY

ICC 2008 Beijing

Department of Information EngineeringUniversity of Padova, ITALY

APOS: Adaptive Parameters Optimization Scheme for Voice over IEEE 802.11g

N. Baldo, F. Maguolo, S. Merlin, A. Zanella, M. Zorzi

D. Melpignano, D. Siorpaes

ICC 2008

Special Interest Group on NEtworking & Telecommunications

Speaker: Michele Zorzi

ICC 2008 Beijing

Motivations (I) VoIP over WiFi is gaining popularity thanks to

the widespread diffusion of WiFi-enabled devices: mobile phones, laptopos,

PDAs VoIP applications: Skype, Yahoo Messenger, MSN

Messenger, Linphone… However, providing good voice quality is a

challenging task due to several factors Medium access contention, radio channel dynamics,

variable loss rate, delay, jitter, etc…

ICC 2008 Beijing

Motivations (II) Some are out of the control of a single STA…

wired part of the e2e connection Round trip delay, packet dropping in congested router,

wireless part of the e2e connection medium contention, radio signal quality, interference level

… while others are under control of the STA application level

voice codec, playout buffer scheme, error concealment techniques,…

MAC/PHY level modulation scheme, retry limit, buffer size, contention

window, …

ICC 2008 Beijing

Aim of the work Goal: providing good quality of VoIP over WiFi under

varying connection conditions Method: Adaptive Parameters Optimization Scheme

Cross-layer iterative adaptation of system parameters Basic idea: estimate, foresee, select

Estimate factors that you cannot control current system state (focused on wireless part only)

Foresee what you would get by acting on the factors that you can control QoS for different (MAC) parameters setting by using a mathematical

model of the system Select the setting that maximizes QoS

Features: non cooperative, standard compliant, modular

ICC 2008 Beijing

APOS architecture

Voice codec

RTP

UDP

IP

LLCMAC 802.11

PHY 802.11

MSE

WLM

APO

Ploss

Te2e

QEB

RQ

Opt

MAC counters

APOS architecture

MSE = Medium State EstimationWLM = Wireless Link ModelQEB = Quality Evaluation Block APO = Adaptive Parameters Optimization

= Fixed parameters vector (Codec, Playout buffer delay) = Tuneable parameters vector (R, rmax) = System state estimate (Ploss, Te2e)

ICC 2008 Beijing

Parameters vectors (, ) Vectors & include all controllable parameters We distinguish between fixed parameters ()

Parameters that hardly change in time voice codec, playout delay, packet aggregation level, wired-path

latency,… …or that cannot be changed preserving std compliancy

contention window, transmission power, … And tuneable parameters: =[rmax,R]

rmax = max retry limit R = PHY transmission rate

We optimize over vector only though APOS can potentially apply to all the controllable parameters!!!

ICC 2008 Beijing

APOS architecture

Voice codec

RTP

UDP

IP

LLCMAC 802.11

PHY 802.11

MSE

WLM

APO

Ploss

Te2e

QEB

RQ

Opt

MAC counters

APOS architecture

MSE = Medium State EstimationWLM = Wireless Link ModelQEB = Quality Evaluation Block APO = Adaptive Parameters Optimization

= Fixed parameters vector (Codec, Playout buffer delay) = Tuneable parameters vector (R, rmax) = System state estimate (Ploss, Te2e)

ICC 2008 Beijing

System state vector () System state vector shall…

summarize the collective effect of factors that cannot be controlled

be almost invariant to single STA parameters setting

We set =[Pcoll, TB, SNR] where Pcoll: Collision probability TB : Channel busy period SNR: Signal to Noise Ratio

MAC factors

PHY factor

ICC 2008 Beijing

Pcoll NRs NRf

N Is NTs NTf

TB CB CBo

NRs NRf

MSE: Medium State Estimate

Pcoll and TB can be directly obtained from MAC counters/measuraments

standard management information base (MIB) NTs: # ack.ed MSDU tx by the STA NTf: # non ack.ed MSDU tx by the STA NRs: # received data frames with valid FCS NRf: # received data frames with invalid FCS

channel occupancy statistics seen by the STA NIs: # of idle slots CB: overall channel busy time CBo: channel busy time due to own tx

The state vector is periodically estimated by the MSE block by using locally available information

ICC 2008 Beijing

MSE: Medium State Estimate SNR can be determined from

MAC counters

+ basic probability properties:

PF NTf

NTs NTf

Psnr PF Pcoll

1 Pcoll

Psn

r

+ pre-computed PHY performance curves

SNRPsnr 1 PF Pcoll

1 Pcoll

ICC 2008 Beijing

APOS architecture

Voice codec

RTP

UDP

IP

LLCMAC 802.11

PHY 802.11

MSE

WLM

APO

Ploss

Te2e

QEB

RQ

Opt

MAC counters

APOS architecture

MSE = Medium State EstimationWLM = Wireless Link ModelQEB = Quality Evaluation Block APO = Adaptive Parameters Optimization

= Fixed parameters vector (Codec, Playout buffer delay) = Tuneable parameters vector (R, rmax) = System state estimate (Ploss, Te2e)

ICC 2008 Beijing

WLM: Wireless Link Model WLM feeds the QEB with the expected packet loss rate (Ploss) and average end to

end delay (Te2e) at the varying of the <> input vectors

We assume that wired-path delay and losses are negligible however, the model can accommodate these factors by using, for instance, RTCP

reports

Te2e is mainly due to the wireless link delay and the playout buffer

Ploss is given by frame dropped by the 802.11 card and frames arriving after their playback time

Assuming that the buffering time buff is given, the WLM outputs depend on the wireless link losses and delay only!

ICC 2008 Beijing

Wireless link delay Goal: estimate mean ms, and variance s

2 of the system time Time taken by a voice packet to cross the wireless link

Method: model the wireless link as a D/G/1 queue-server system where customers MPDUs generated by the upper layers arrival rate frame-generation rate of the voice codec (fixed) server MAC entity service time y time taken by the MAC entity to process a MPDU (either delivering it

to the peer unit or dropping it after rmax unsuccessful tx attempts)

We assume {yj} are i.i.d r.v. also independent of the arrival process

Complete statistical analysis is available in the literature [Servi-’86] but requires roots search for complex polynomial! not suitable for portable devices (as cel. phones)

We prefer to relinquish the complete statistic in favour of a simpler (though less accurate) estimate of first and second order delay statistics…

ICC 2008 Beijing

Service time statistics (1/2) Given [Pcoll,TB,SNR] , [R,rmax] and [L,...] the service time is given by

Working this expression we get* the 1st, 2nd and 3rd-order moments of y my=E[y] My=E[y2] My(3)=E[y3]

* See [Zanella, De Pellegrini, ComLet05]

# failures before success

y s i j,hj1

rh

h0

i

iTF

i0

rmax 1

TS 1 d d rmaxTF

h-th backoff stage

Failed tx time

Successful tx time

s(i)=1 if success after i failures, s(i)=0 otherwise

d=1 if Pck drop

Failed tx time

ICC 2008 Beijing

Service time statistics (2/2)just a flavour of the results (details in the paper)

ICC 2008 Beijing

Wireless link delay System time s can be expressed as

x: number of customers (MPDUs) in the system at the arrival epoch y’: residual service time of the (possible) customer in service at the arrival epoch H(x) : Heaviside function

taking expectations we finally get statistical mean (ms):

statistical power (Ms):

= my = Pr[x > 0] (load factor)

whereas, applying renewal theorem, we get:

sy y y jj1

x 1

H x

ms my m y mx my

Ms My 1 M y mx 2mym y y2 Mxmy

2

m y My

2my

; M y M y 3

3my

ICC 2008 Beijing

Packet losses Losses due to the wireless link

Losses due to playout underrun Pbuff & delay jitter s are related through the Chebyshev bound*

Hence, Pbuff can be estimated as

Overall losses

2

22,1minbuff

sbuffP

Pdrop PFrmax

Ploss Pdrop 1 Pdrop Pbuff

22 2 sbuffbuff P

ICC 2008 Beijing

APOS architecture

Voice codec

RTP

UDP

IP

LLCMAC 802.11

PHY 802.11

MSE

WLM

APO

Ploss

Te2e

QEB

RQ

Opt

MAC counters

APOS architecture

MSE = Medium State EstimationWLM = Wireless Link ModelQEB = Quality Evaluation Block APO = Adaptive Parameters Optimization

= Fixed parameters vector (Codec, Playout buffer delay) = Tuneable parameters vector (R, rmax) = System state estimate (Ploss, Te2e)

ICC 2008 Beijing

QEB: Quality Evaluation Block

The QEB implements the utility function Q proposed in [Boutr.03] inspired by the E-Model ITU-T recommendations

G.107 and G.113

R=Q(vcodec, Ploss,Te2e)

R can be mapped to Mean Opinion Score (MOS)

[Boutr.03] C. Boutremans and J.-Y. Le Boudec, “Adaptive joint playout buffer and fec adjustement

fot internet telephony,” in INFOCOM 2003

Plo

ss

MOS countours

Te2e [ms]

ICC 2008 Beijing

APOS architecture

Voice codec

RTP

UDP

IP

LLCMAC 802.11

PHY 802.11

MSE

WLM

APO

Ploss

Te2e

QEB

RQ

Opt

MAC counters

APOS architecture

MSE = Medium State EstimationWLM = Wireless Link ModelQEB = Quality Evaluation Block APO = Adaptive Parameters Optimization

= Fixed parameters vector (Codec, Playout buffer delay) = Tuneable parameters vector (R, rmax) = System state estimate (Ploss, Te2e)

ICC 2008 Beijing

APO: Adaptive Parameters Optimization

do forever MOSopt=0 get =[Pcoll, TB, SNR] from MSE for =[rmax,R] in parameter space do

compute [Ploss,Te2e]=WLM() compute MOS = Q(Ploss,Te2e) If MOS > MOSopt,

• MOSopt = MOS opt =

endif endfor

endo

ICC 2008 Beijing

Model Validation

Estim

ated

Pco

ll and

Per

r

SimR =54 MbpsSNR = 20 dB

P loss

SNR=14 dB R =18 Mbps

SNR=30 dB R =54 Mbps

contending STAs

ModSim

T e2e [

ms]

SNR=14 dB R =18 Mbps

SNR=30 dB R =54 Mbps

contending STAs

ModSim

contending STAs

MSE: Medium State Estimator validation

WLM: Wireless Link Model validation

ICC 2008 Beijing

Results Theoretically optimal

setting in the space TB fixed

Optimal R and rmax depend on both Pcoll, with low Pcoll, rate

adaptation can be more aggressive

with high SNR, rmax strongly depend on Pcoll

ICC 2008 Beijing

Results Parameters optimization

permit to maintain good MOS in many situations Red lines: std param.s setting Black lines: APOS

However, estimation & model errors may penalize the performance gain especially with few users

ICC 2008 Beijing

Conclusions APOS framework is a stand-alone, std compliant solution

mathematical model based on of network status information locally available in commercial devices

APOS enhances VoIP performance by cycling over three steps 1) estimate the contention level and quality of the wireless medium 2) model the effect of parameters tuning on e2e QoS 3) select the best configuration for the current medium status

Optimization can be performed on any subset of tuneable parameters We reported only R & rmax optimization, but we also tested playout buffer, voice

codec and so on... Although parameters might be tuned independently one another (as done

by today rate adaptation algorithms), joint optimizing of multiple parameters yield much better results!

APOS scheme is specialized to voice application, though the general framework might be adapted to other QoS services