Enabling Multimedia QoS Control with Blackbox Modelling

1
Enabling Multimedia QoS Control with Black-box
Modelling
Gianluca Bontempi IMEC -DESICS/MICS   Leuven,
Belgium
Gianluca Bontempi Université Libre de
Bruxelles Brussels, Belgium gbonte_at_ulb.ac.be
Gauthier Lafruit IMEC-DESICS/MICS   Leuven,
Belgium lafruit_at_imec.be
2
Multimedia applications
functionality constraints
architecture
time/energy constraints
quality constraints
user
3
Outline

• Quality of Service (QoS) and multimedia.
• A Control Interpretation of the QoS Problem.
• A Data Analysis Procedure for QoS Modeling.
• The VTC/MPEG-4 Modeling Problem.
• The Experimental Results.
• Conclusions and future work.

4
QoS and multimedia

• Quality of Service (QoS) methods aim at trading
  quality vs. resources to meet the constraints
  dictated by the user, the functionality and the
  platform.
• QoS originally developed in network
  communication.
• QoS recently extended to the domain of multimedia
  processing.
• QoS relevant in multimedia scalable systems,
  where the resources and the functionality can be
  controlled by a set of parameters.

5
The QoS dilemma
Find the optimal balance (within the systems
limitations)
6
Scalable systems
SYSTEM
7
(No Transcript)
8
Dynamic Load Variations
Factor 6
9
QoS in multimedia and control

• Multiple criteria to be optimised.
• Control parameters in scalable systems.
• Non stationary environments.
• Complex relations between control parameters and
  criteria.

10
QoS control
user preferences
quality
control parameters
MULTIMEDIA ALGORITHM
performance
QoS CONTROLLER
11
Supervised learning
input
MULTIMEDIA ALGORITHM
output
error
OBSERVATIONS
MODEL
prediction

• Finite amount of noisy observations.

• No a priori knowledge of the input/output
  relation.

12
Predictions with Lazy Learning
query
query
query
13
Awards in international competitions

• Data analysis competition awarded as a runner-up
  among 21 participants at the 1999 CoIL
  International Competition on Protecting rivers
  and streams by monitoring chemical
  concentrations and algae communities.
• Time series competition ranked second among 17
  participants to the International Competition on
  Time Series organized by the International
  Workshop on Advanced Black-box techniques for
  nonlinear modeling in Leuven, Belgium

14
Industrial Applications

• Financial prediction of stock markets in
  collaboration with Masterfood, Belgium.
• Prediction of yearly sales in collaboration with
  Dieteren, Belgium, the first Belgian car dealer.
• Non linear control and identification task in
  the framework of the Esprit project FAMIMO.
• Modeling of industrial processes in
  collaboration with FaFer Usinor steel company in
  Belgium, and Honeywell Technology Center, US.
• Performance modeling of embedded software in
  collaboration with Philips Research.

15
The problem

• Predicting the resource requirements of the VTC
  wavelet-based algorithm of a MPEG-4 decoder as a
  function of the value of some control parameters
  on the basis of a finite amount of data.

16
Experimental setting

• Benchmarks.
• Control parameter selection.
• Predicted variables.
• Data collection (Atomium).
• Model calibration
• linear
• non linear (Lazy Learning)
• Validation
• training and test
• Results

17
Benchmarks

• 21 image test files in yuv format.
• Lena picture
• images from 4 different AVI videos
• Akiyo, IMECnology, Mars, Mother and Daugthter.
• Software MoMuSys (Mobile Multimedia Systems).
• Microprocessor HP J7000/4 at 440 MHz.

18
Input parameters

• w Image width.
• h Image height.
• l Number of wavelet levels. I ?2..4
• q Quantization type (SQ, MQ or BQ).
• n Target SNR level. n ? 1..3.
• y Quantization level QDC_y. y ?1, 6.
• u Quantization level QDC_uv. u ?1,6.
• ee Encoding execution time.
• re Total number of encoding read memory
  accesses.
• we Total number of encoding write memory
  accesses.

19
Predicted variables

• Execution time.
• Number of memory reads (obtained by using the
  IMEC Atomium tool).
• Number of memory writes (obtained by using the
  IMEC Atomium tool).

20
Input dataset
21
Input/output dataset
22
Cross-validation
All samples takes part once to the test set.
23
Results
PE percentage error.
24
(No Transcript)
25
Future work

• extending the work to multiple platforms and
  architecture,
• exploring prediction models of some quantitative
  attributes of the quality,
• integrating the prediction models in a control
  architecture, negotiating online the quality
  demands vs. the resource constraints,
• integrating the application-level control with an
  higher system-level control mechanism (e.g.,
  resource manager).

26
References

• Atomium http//www.imec.be/atomium/
• Lazy Learning http//iridia.ulb.ac.be/lazy/