Sie sind auf Seite 1von 23

Euro NF Ph.D.

Euro-NF Ph D Course on QoE

-Quantitative Relationships Between QoE and QoS


Markus Fiedler

How to derive QoEQoS QoE QoS relationships?


Experiments
(Groups of) {nave|expert} user(s) Well-defined conditions and network stimuli

Data collection
Subjective: questionnaires, observations Objective: j Measurements

Analysis
Trends, Trends relationships, relationships thresholds Regressions, hypotheses, impact analysis

Dreamhack Winter 2007 Gaming QoE

Dreamhack Winter 2007 Shaping / QoS measurements

QoE in ITU-T ITU T Rec. Rec G.1080 G 1080


Minimum level of transport layer performance required to provide satisfactory QoE for H.262 encoded HDTV services
Jitter < 50 ms Max. duration of single error event: 16 ms
Corresponding loss period in IP packets: < 24 @ 15 15.0 0 Mbit/s < 29 @ 18 18.1 1 Mbit/s

Loss distance: 1 error event per 4 hours


Corresponding average IP video stream packet loss rate: 1.17106

Mobile videos with loss introduced


MOS vs. loss ratio
5 QoEMoVi PhD students + supervisors Dreamhack, > 20 years D Dreamhack, h k < 20 years Summer school 2 BTH students avg 1 0.0% 0.2% 0.4% 0.6% 0.8% 1.0%

Fundamental QoEQoS QoE QoS relationships


Dependencies on network conditions typically addressed by parameter vectors [QoSi, , QoE]
Results from q questionnaires, , observations, , measurements Several impact factors: QoE = f(QoS1, QoS2, )

We focus on one impact factor at a time Description by partial differential equations Consider fundamental relationships p of the type yp
QoE = g (QoE , QoSi ) QoSi

Maximise/minimise QoE to interval [1, 5] afterwards

Fundamental QoEQoS QoE QoS relationships


Investigated set
Linear Logarithmic Exponential Power QoE QoSi QoE log(QoSi) log(QoE) QoSi log( g(Q QoE) log( g(Q QoSi)

Nice properties
Seen from regressions on linear vs. logarithmic scales Reasoning behind each relationship

Examples
Most formulae from [Shaikh et al., 2010] Work supported by Euro-NF

Linear relationship
QoE axis Linear scale Additive change

QoE QoSi
QoS axis Linear scale Additive change

QoE gradient independent of QoE and QoS Linear regression often the first choice Local approximation Example:
Download time perception as function of loss

QoE 4.3 31PLR (2 > 0.99)

Logarithmic relationship
QoE axis Linear scale Additive change

QoSi QoE QoSi

QoS axis Logarithmic scale Multiplicative change

QoE gradient proportional to reciprocal QoS Weber-Fechner Law (1834):


Just noticeable differences, multiplication on stimuli side

Utility functions (implicit proportional fairness) Example:


Download time perception as function of bandwidth

QoE 1.2 1 2+3 3.3lg( 3lg(R/Mbps) (2 > 0 0.99) 99)

Exponential relationship
QoE axis Logarithmic scale Multiplicative change

QoE QoSi QoE

QoS axis Linear scale Additive change

QoE gradient proportional to actual QoE


Nuclear decay Human memory IQX hypothesis [Hossfeld et al., 2007; Fiedler et al., 2010]

Examples
Image quality perception as function of blur, blockiness, (QoP) Download time perception as function of response time (QoD)

QoE 4.8exp(0.15RT /s) (2 > 0.99)

Power type relationship Power-type


QoE Q E axis i Logarithmic scale Multiplicative change Long tails on both axes Examples
Session volume as function of bandwidth Video perception as function of jitter
6 5 4
OS

QoE QoSi QoE QoSi

QoS Q S axis i Logarithmic scale Multiplicative change

0.39 2 QoE 14.6(PDV /ms) ( 0.68) OS vs PDV (ms)

y = 14.613x-0.386 R = 0.6803

3 2 1 0 0 100 200 300 400 PDV(ms) 500 600 700 800

Provisioning delivery hysteresis Provisioning-delivery


Different types of QoE and QoS parameter Distinguish between
Success and failure Resource and problem Promise-making and promise-keeping

Examples p Provisioning-delivery hysteresis


Derivation Examples

Different types of QoE and QoS parameters


Success rating QoE Q E
The higher, the better Mean Opinion Score (1 (1..5) 5)

Resource measure

Q Sr QoS

The higher, the better Throughput

Success measure Failure rating QoE


The higher, the worse Cancellation rate Churn rate

QoSs

Availability (e.g. (e g 99.99 99 99 %) Packet success ratio

Failure measure

QoSf

Watch the signs!

The higher, the worse P k tl Packet loss ratio ti Delay jitter Reordering

Skype: MOS = f(packet loss ratio)


[Fiedler et al., 2010] [ ]

QoE

QoSf

Skype: MOS = f(reordered ratio)


[Fiedler et al., [ , 2010] ]

QoE

QoSf

G 1030/download: MOS = f(session time) G.1030/download:


[Fiedler et al., [ , 2010] ]

QoE

QoSf

Web: Cancel-rate Cancel rate = f(delivery bandwidth)


[Khirman & Henriksen, 2002] [Fiedler et al., [ , 2010] ]

QoE

QoS r

Web: MOS = f(throughput)


[Shaikh et al., 2010] [ ]

QoE

QoS r

Basic shapes
Q E QoE
resource
F L I P

QoE
FLIP

QoE
problem

QoSr

QoSf

QoSs

QoE

Relationship? p

QoS r

Provisioning delivery hysteresis for web Provisioning-delivery

1 PLR [Fi dl & Hossfeld, [Fiedler H f ld 2010]

Provisioning delivery hysteresis for video Provisioning-delivery


Potential for savings

To be avoided

[Zi [Zinner et t al., l 2010]

Conclusion
Given you are to study QoEQoS relationships
Get users involved its fun! Try objective measurements of user action
Check out [ETSI STF 354]

Play y with the logscale g button on both axes


Eventually, offsets need to be taken away

Interprete the outcome


Additive Additi or multiplicative lti li ti change? h ? Implications? I li ti ?

Please let me know of new fundamental relationships and examples you publish: markus.fiedler@bth.se

Das könnte Ihnen auch gefallen