Transcript Slides - TERENA Networking Conference 2004

Intelligent and Adaptive Middleware to
Improve User-Perceived QoS in Multimedia
Applications
Pedro M. Ruiz, Juan A. Botia,
Antonio Gomez-Skarmeta
University of Murcia
Terena Networking Conference 2004
Rhodes, June 2004
Drivers for adaptive applications
• E2E QoS requires local resource management
• Terminals are heterogeneous and media adaptation is
needed
• Network conditions are unpredictably changing and not
under control (e.g. Ad hoc nets, PLC networks, etc.)
• QoS in terms of bandwidth and delay cannot be
guaranteed just with network-layer QoS mechanisms
• In these cases, user-perceived QoS can be improved
using applications being able to adapt to:
–
–
–
–
Network conditions
QoS Violations
Shortage of local resources (eg. CPU, Memory, etc.)
User preferences
2
Adaptive Applications
3
Video Quality vs. Bandwidth
10 FPS, SQCIF both
for MJPEG and
H.263
1190 Kbps
300 Kbps
210 Kbps
140 Kbps
Non-linear
Perception
70 Kbps
50 Kbps
30 Kbps
4
Towards user-awareness...
• Traditional approaches based on profiles
– Simple and easy to implement
– Usually are not fine-grained enough
– Are not able to capture the perceptual preferences
• Adaptations based on low level parameters (e.g.
Bandwidth, packet losses, etc)
– Do not really consider user preferences
– Perceptual QoS is not linearly related to those low level metrics
• Previous works focused on evaluating the impact of
each parameter on the user perception
– There is not a real model of the user, particularly when multiple
parameters can be tuned simultaneously (e.g. codecs, frame
rates, etc.)
5
Architecture for Multimedia
Adaptive Applications
Type
Seq #
Loss %
Delay
User pref
Estimated BW
App. Layer
Audio Codec
Video Codec
Frame Rate
Size
Quantization
...
App. Layer
QoS signaling
QoS signaling
Adaptation
Logic
Net. Layer Mon
Net.Layer Mon
Audio
...
Video
Slides
Audio
Adaptation
Logic
Video
TCP/UDP
TCP/UDP
IP
IP
MAC
MAC
PHY
PHY
...
Slides
6
Basic Adaptation Logic
START
QoS Report
lost++
no
QoS Report?
yes
yes
Loss > 5%
QoS Report Lost
>3
yes
no
Loss = 0%?
no
Reset Zero-loss
count
yes
Zero-loss++
no
Zero-loss > 15
no
yes
Downgrade Quality
Upgrade Quality
END
7
Why applications aware of the
user-perceived QoS?
• There are many ways to adapt data-rates to the available bandwidth
–
–
–
–
–
–
–
Audio & Video Codecs
Video Quantization factor
Audio sampling rate
Video frame rate
Video size
Component selection
Buffering
Which combination
would be preferred by
the user?
• Not trivial to select a new combination of settings satisfying the users
– Reduce frame size?, reduce frame rate?, change codec?
• Traditional adaptive applications improve user-perceived QoS but
they offer sub-optimal solutions
• Adaptive applications should be able to deal with the
user perception of QoS!
8
Our proposal for user awareness
• Use of machine learning techniques to help at
modeling the user perceived QoS
– Number of possible combinations of application
settings is big enough!
– Perceptual QoS may change from one indivudual to
another and it is extemely complex to be analitically
modelled
– A “black box” model may resemble the usersatisfaction without needing to understand the
complex processes involved in user perception
9
Modelling user-perceived QoS
• Difficult to model, due to the subjective aspects
involved
• We apply a rule induction machine learning
algorithm over learned data
Initial data-set
(864 entries)
BW
ACOD
VCOD
FSIZE
Quant
FPS
Loss%
Score
33, 56, 88, 128, 384Kb/s
PCM, G.711u, G722, GSM
MJPEG, H263
CIF, QCIF, 160x128
5, 10, 15, 30, 60
2..24
0..100%
1..5 (according to MOS)
SLIPPER algorithm with t=5
Set of rules representing
user-perceived QoS
10
Rules Generated by SLIPPER
Some of the lessons
learnt from rules
Higher FR => higher QoS
but user’s prefer lower FR
(not below 4 FPS) provided
that the video is bigger
In most cases PCM audio
is not required. The
bandwidth savings can be
used to improve other
components
Audio QoS has greater
impact
Etc..
if matchConfidence {
[QFVIDEO >= 60, VIDCOD = MJPEG,
FSIZE = QCIF, LOSS <= 10, FPS >= 6] -> 2.8792
[AUDCOD = GSM, BW >= 80,
QFVIDEO >= 30, FSIZE = QCIF, FPS <= 6] -> 1.4357
[AUDCOD = GSM, BW >= 128, LOSS = 0,
QFVIDEO >= 30, FPS >= 3, VIDCOD = MJPEG] -> 1.7013
[] -> -2.4188
} > 0 then 5 else if matchConfidence {
[BW >= 384, QFVIDEO >= 40, FSIZE <= 2] -> 2.7121
[QFVIDEO >= 30, VIDCOD = MJPEG,
LOSS <= 3, AUDCOD = G722, BW >=80] -> 1.1756
[FSIZE = CIF, QFVIDEO >= 30,
LOSS <= 3, AUDCOD = G722, BW >= 80] -> 1.4437
[] -> -1.5044
} > 0 then 4 else if matchConfidence {
[LOSS >= 30] -> 2.1188
[QFVIDEO <= 5] -> 1.4142
[LOSS >= 16, FPS <= 3] -> 1.5438
[] -> -1.0984207275826066
} > 0 then 1 else if matchConfidence {
[LOSS >= 16] -> 1.9109
[QFVIDEO <= 10, FSIZE = QCIF] -> 1.5861
[FSIZE = 160X128, QFVIDEO <= 40, VIDCOD = H.263] -> 1.2546
[] -> -0.3953
} > 0 then 2 else 3
11
Empirical Results
• Scenario
– Real MMARP-based ad
hoc network
– Path specifically selected
to guarantee variability
• Application
– ISABEL-Lite with
extensions
• Trials
– Traditional multimedia
application
– Adaptive multimedia
application
12
Total Losses
13
Histogram audio/video loss-rate
14
User’s Mean Opinion Scores
15
Conclusions and Future Work
• Adaptive applications have demonstrated to be effective
in wireless and mobile scenarios
• The machine learning user’s modelling has shown to be
effective
• Applications aware to the user-perceived QoS have
demonstrated to offer to better satisfy user’s QoS
expectations in a real ad hoc wireless networks
• Optimization on the triggering of the adaptation have
demonstrated Future work include among others
– Reinforcement learning inside the terminal
– Combination with user profiling mechanisms
16
Intelligent and Adaptive Middleware to
Improve User-Perceived QoS in Multimedia
Applications
Pedro M. Ruiz, Juan A. Botia,
Antonio Gomez-Skarmeta
University of Murcia
Terena Networking Conference 2004
Rhodes, June 2004