Upload
nguyenxuyen
View
215
Download
0
Embed Size (px)
Citation preview
Error-Aware Video Encoding to Extend EnergyQoS
Tradeoffs for Mobile Embedded Systems1
KYOUNGWOO LEE
University of California, Irvine
NIKIL DUTT
University of California, Irvine
and
NALINI VENKATASUBRAMANIAN
University of California, Irvine
(1) Problem domain –
(2) Specific motivation and problem –
(3) Proposal –
(4) Contributions –
Categories and Subject Descriptors: []: —;
General Terms:
Additional Key Words and Phrases:
1. INTRODUCTION
(1) Emerging video applications on mobile devices and problem domain—video applications are becoming more popular.—energy efficiency is a key property to consider for desinging mobile embedded
systems.
1This is an expanded version of a paper published in the Proceedings of the IFIP WorkingConference on Distributed and Parallel Embedded Systems (DIPES) 2008. The currentmanuscript extends the previous paper by exploring the design space for the partially protectedinstruction caches and proposing the heuristic algorithms to efficiently find the best configurationwith minimal vulnerability under the performance penalty.
Author’s address: K. Lee, N. Dutt, and N. Venkatasubramanian, Department of ComputerScience, University of California, Irvine, CA 92697.
Permission to make digital/hard copy of all or part of this material without fee for personalor classroom use provided that the copies are not made or distributed for profit or commercialadvantage, the ACM copyright/server notice, the title of the publication, and its date appear, andnotice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish,to post on servers, or to redistribute to lists requires prior specific permission and/or a fee.c© 20YY ACM ????-????/20YY/0?00-0001 $5.00
(2) Previous Work
—Energy-aware video applications for mobile embedded systems
—Still, energy-efficient approaches are required to prolong the life of mobiledevices.
(3) Problem Definition
—wireless network is not stable.
—error-resilient video encoding is important.
—some error-resilient video encoding is energy-efficient as well.
—what if network is stable (i.e., no errors) but error-resilient video encodingwill work as a normal or standard video encoder.
(4) Our approach – Error-Aware Video Encoding (EA-VE)
—Intentional error injection (e.g., frame dropping) can be an effective knob totradeoff QoS for the energy efficiency.
—QoS-feedbacked mechanism can maintain the video quality in EA-VE schemes.
(5) Results and contributions
—Save the huge amount of energy consumption at the minimal cost of the videoquality.
—Our approaches can extend the design space significantly to tradeoff the videoquality for the energy reduction or vice versa.
2. THE BACKGROUND
2.1 Energy/QoS-aware Video Encoding
—Overview of quality aware viceo encoding
—Overview of energy aware video encoding
VLC : Variable Length CodingQ : QuantizationDCT : Discrete Cosine TransformME : Motion Estimation
ME DCT Q VLC
A
B
C
DB
C
A
D Error Exploitation(e.g., error injection)
Energy/QoS−awareVideo Encoding(Section 2.1)
Error−Resilient
(Section 2.2)
Standard Video Encoder
Constraint
Our Proposal
Quality of Service(e.g., PSNR)
Energy−Efficiency
Error−Resilience
Quality Control(e.g., Resolution, IP Ratio,
and Quantization [15])
Power Management
Resilience Control
i
i
ii
i ii iii
i ii iii
i
ii
iii
(e.g., Voltage Scaling [14,25])
(e.g., Intra−MB Threshold [10])
Video Encoding
(e.g., 10% packet loss rate)
(e.g., Residual Power)
[10,14,15,25]: References
i,ii,iiiA,B,C,D : Knobs
Knob
: Constraints
Energy−EfficientVideo Encoder
Error−ResilientVideo Encoder
Error−AwareVideo Encoder
Fig. 1. Constraints and Knobs considered by Previous Approaches and Our Pro-posal
2.2 Error-Resilient Video Encoding
—Overview of robust video encoding techniques
—Promising Intra Refresh encoding techniques
2.3 Error-Aware Video Encoding: Our Proposal
—passive error-aware video encoding
—error-aware energy reduction at network
—active error exploitation – our proposal
3. OUR APPROACH
3.1 System Model
(1) Our system model—Mobile Video Conferencing
IIFDT
WNI
Tx
WNI
Rx IIIFDT
Dec
CPU
IFDT
AP : Access PointWAN : Wide Area Network
: Video Data Flow
Rx : Receiver: TransmitterTx
FDT : Frame Drop TypeEnc : EncoderDec : Decoder
CPU : Central Processing UnitWNI : Wireless Network Interface
wiredwiredWAN
Network
AP 1 AP 2
wireless wireless
Mobile 1 Mobile 2
Transmission Errors
CPU
Enc
Fig. 2. System Model (Mobile Video Conferencing) and Frame Drop Types I/II/IIIfor Active Error Exploitation
—CPU and WNI are dominant contributors to power consumption—Why one path from an encoder to a decoder
3.2 Fundamentals of Active Error Exploitation
(1) Intentional error injection—Network is unreliable—ER and EC are designed against network errors—Error injection such as frame skipping can occur everywhere—Error injection can affect the following components in terms of EC and QoS—We study an active error exploitation at the Encoder since it is the most
effective
(2) Motivated examples—Error-aware video encoding techniques can present larger tradeoff spaces
compared to a normal video encoder (GOP-15) and an original PBPAIR—GOP-15 is a base video encoder for comparison, which indicates 15 P-frames
between two I-Frames.—PBPAIR is a previously proposed error-resilient video encoding technique—EA-PBPAIR presents a novel knob, error-injection rate (EIR), which is able
to tradeoff the video quality and the energy consumption
Table I. Energy Consumption Category
Type Description
Enc EC Energy consumed by CPU to encode a video streamTx EC Energy consumed by WNI to transmit an encoded video streamSrc EC Enc EC + Tx EC
Dec EC Energy consumed by CPU to decode a received video streamRx EC Energy consumed by WNI to receive a video streamDst EC Dec EC + Rx EC
EC = Energy Consumption
Encoding Energy Saving (FOREMAN 30 frames)
30
35
40
45
50
0 1000 2000 3000
All Possible Frame Loss Patterns (10% EIR)
En
erg
y S
av
ing
Ra
te (
%)
(a) Energy Reduction for En-coding
Decoding Energy Saving (FOREMAN 30 frames)
0
5
10
15
20
0 1000 2000 3000
All Possible Frame Loss Patterns (10% EIR)
En
erg
y S
av
ing
Ra
te (
%)
(b) Energy Reduction for De-coding
PSNR Degradation (FOREMAN 30 frames)
0
2.5
5
7.5
10
0 1000 2000 3000
All Possible Frame Loss Patterns (10% EIR)
PS
NR
De
gra
da
tio
n R
ate
(%
)
(c) Quality Degradation
Fig. 3. Energy Consumption Saving and Quality Degradation of Error-Aware VideoEncoder (EA-PBPAIR) compared to a standard video encoder (GOP-15)
3.3 EA-VE: Error-Aware Video Encoding
Error-InjectedVideo Data
OriginalDataVideo
Error-AwareVideo Data
Feedback(e.g., Frame Dropping)
Error Controller Error-ResilientVideo Encoder
(e.g., PBPAIR)Parameters
Error-Injection Unit Error-Canceling Unit
Constraints(e.g., quality requirement)
(e.g., quality feedback,packet loss rate)
(e.g., Error Rate,Intra Threshold)
Knob(e.g., error injection rate)
Error-Aware Video Encoder
Fig. 4. Error-Aware Video Encoder consists of Error-Injection Unit and Error-Canceling Unit.
(1) Introduction and wrap-up—Intentional error-injection can help VCS save the energy consumption
(2) Two-steps error-aware video encoder—Error-aware video encoding technique we present is composed of two units
such as error-injection unit and error-canceling unit—We present two simple approaches, which exploit the error-tolerance of video
data to increase the energy reduction for mobile video applications—(i) Error Controller
—Error-injection unit preprocesses the parameters for the following video en-coder, and inserts errors purposely to tradeoff between energy and qualityof service
—(ii) Error-Resilient Video Encoder—Error-canceling unit can be a normal video encoder or an error-resilient
video encoder.
—They reduce the effects of error-injection thanks to the nature of error-tolerance of video data and/or thanks to the error-resilience.
(3) In conclusion, EA-PBPAIR can save the energy consumption
—(i)—(ii)—(iii)—(iv)
3.3.1 Frame Dropping
(1) Simple error injection scheme is to skip the frames periodically
—Intelligent frame skipping can increase the quality—Different strategy from general frame skipping technique—Any frames can be dropped in EA-VE—For our approaches, we use “Frame Dropping Technique” at the error-injection
unit, and “Error Resilitne Video Encoder” at the error-canceling unit
(2) PFD
(3) MDFD
3.3.2 Error Canceling
(1) PBPAIR is energy efficient
—PBPAIR is a video encoding technique which is not only error-resilient butalso energy-efficient.
—PBPAIR has two parameters such as Intra Th and PLR (Packet Loss Rate)
(2) GOP
(3) PGOP
3.4 Adaptive EA-VE
QcQc
f1f2
f3
f0Original Video Data
f3
f0f2
Error−Injected Video Data
Increase EIR Decrease EIR
QfQc >
������������������������������
������������������������������
������������������������������
������������������������������
Lossy Video Dataf0
f2
EIRI
Qf
������������������������������
������������������������������
������������������������
������������������������
������������������������������
������������������������������
f3
f0f2
Error−Aware Video Data
para 2para 1 PFD
: input (constraint and parameter): feedback: control: video data
IEIR I
Error Injection (e.g. PFD)
Update EIR
NO
YES
Adaptive EIR
BEGIN
Error Exploiting Video Encoder
Error Controller
Dec
oder
f1 is dropped
f3 is lost
Error−Resilient Video Encoder
Lossy Networkpara = EIR + PLR + AER1 PLR
AER
: Periodic Frame Dropping
fn : n th video frame
EIR = EIR
CONSTRAINT)
INJECTION RATE)(Initial ERROR
(QUALITY
(QUALITYFEEDBACK)
(ADJUSTEDERROR RATE)
(PACKETLOSS RATE)
para : encoding parameter
Fig. 5. Flowchart of Error Controller for Adaptive EA-VE
(1) We introduce the EIR as a new knob to tradeoff energy and QoS—It opens and expands the operating points, which are newly discovered.
(2) First approach—This approach is to inject errors by dropping frames given an EIR—The sum of an actual network PLR and EIR is an input as Para1 for PBPAIR—For instance, the actual PLR is 10%, and EIR is 5%, then the parameter
PLR to PBPAIR is 15%
(3) Second approach—To adapt EA-PBPAIR for meeting QoS requirement,—EIR can be adjusted for dynamic network status—Change EIR based on the quality feedback from the decoding
(4) Third approach
—AER (Adjusted Error Rate) can be updated to increase the energy saving orto increase the video quality.
4. EXPERIMENTAL SETUP
Packet TrafficFrame Size
Analysis 1 Analysis 2
WNI Energy for Transmit
Analysis 3
CPU Energy for Decoding
Video Stream
CPU Power Parameters PLR
NS2 Simulator
Mobile 1
Encoder Tx
PrototypeSystem
Mobile 2
Decoder
SystemPrototype
Exe Time
Wireless Network Wireless Network
Quality Constraint
Encoding Parameters
Frame Dropping Policy
Error Injection Rate
Frame Dropping TrafficGenerator
Frame-Dropped Pattern
CPU Energy for Encoding(Enc EC)
WNI Energy for Receive(Tx EC)
(Rx EC)Video Quality(Dec EC)
(PSNR)
WNI Power Parameters
Network Topology
Packet LossesArrival Time
Wired NetworkAP 1 AP 2
: Control Flowof Simulation
: Video Data Flow
simulation output
simulation input
operation
temporary data
Rx
Fig. 6. Experimental Framework - System Prototype + NS2 Simulator
(1) Experimental Framework—Integration of System Prototype and NS2 using Perl
(2) System Prototype
(3) Power Values and Parameters for CPU and WNI
(4) Applications
(5) Test sequences such as Akiyo, Foreman, and Coastguard
(6) NS2 Simulator
Table II. Power Parameters
CPU WNI
Power Mode Active Idle Sleep Transmit Receive Idle Sleep
Power (W) 0.411 0.121 0.001 1.425 0.925 0.80 0.045
Transition (msec) 1.00 0.75
Energy Consumption at Source
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Enc Tx Enc + Tx
No
rma
lize
d R
ate
of
EA
-PB
PA
IR t
o G
OP
-8
EA-PBPAIR GOP-8
(a) Energy Consumption at the Source
Energy Consumption and QoS at Destination
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Dec Rx Dec + Rx Quality
No
rma
lize
d R
ate
of
EA
-PB
PA
IR t
o G
OP
-8
EA-PBPAIR GOP-8
(b) Energy Consumption at the Destination
Fig. 7. Energy Reduction and Quality Degradation of EA-PBPAIR compared toGOP-8 (EIR = 10%, PLR = 5%, FOREMAN 300 frames)
Energy Consumption and Quality
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Enc Tx Enc + Tx Dec Rx Dec + Rx Quality
No
rma
lize
d R
ate
of
EA
-PB
PA
IR t
o G
OP
-15
EA-PBPAIR GOP-15
(a) PLR = 0%
Energy Consumption and Quality
0
0.2
0.4
0.6
0.8
1
Enc EC Tx EC Enc EC +
Tx EC
Dec EC Rx EC Dec EC +
Rx EC
Quality in
PSNR
No
rma
lize
d R
ate
of
EA
-PB
PA
IR t
o G
OP
-3
GOP-3 EA-PBPAIR
(b) PLR = 10%
Fig. 8. Energy Reduction and Quality Degradation of EA-PBPAIR compared toGOP-K (EIR = 10%, FOREMAN 300 frames)
5. EXPERIMENTAL RESULTS
5.1 Energy Reduction from Active Error Exploitation
—A set of experiments to show the effects of our error-exploiting on energy reduc-tion without significant quality loss
(1) (A1) The first set of experiments : EA-PBPAIR vs. GOP300 with respect toenergy consumption and quality of service at actual PLR = 0, 5, and 10%,respectively—keep the transmitted data size the same—or include the energy consumption for larger data transmission in EA-PBPAIR
(2) (A2) EA-PGOP vs. PGOP w.r.t. energy consumption and quality of service—emphasize the effects of error-injection on energy consumption at PLR =
10%—clearly show that EA works for an error-resilient video encoding such as
PGOP
Energy Consumption and Quality
0
0.2
0.4
0.6
0.8
1
1.2
Enc EC Tx EC Enc EC +
Tx EC
Dec EC Rx EC Dec EC +
Rx EC
Quality in
PSNRN
orm
ali
ze
d R
ati
o o
f E
A-P
GO
P t
o P
GO
P-
PGOP-3 EAPGOP-4 EAPGOP-6
Fig. 9. Energy Reduction and Quality Degradation of EA-PGOP compared toPGOP (PLR = 10%, EIR = 10%, AER = 0% and -10%, FOREMAN 300 frames)
—Note that it shows the energy reduction at encoding and decoding whileshowing the energy overhead for transmission and receiving the video datasince PGOP does not have a mechanism to adjust the video quality andthe encoding file size as PBPAIR has (e.g., Intra Threshold). However, it isthe reason why we present AER, which may increase the amount of errorswe injected intentionaly to save the energy consumption, especially for thecommunication energy until the quality is satisfied with our feedback-basedmechanism.
—present the quality loss, which is why we propose the feedback-based errorcontrol approach in the next section
(3) (A3) EA-GOP vs. GOP w.r.t. energy consumption and QoS—we apply our active error exploitation for a standard video encoding such
as GOP, where we can see the positive impact of error-exploitation on theenergy saving at the slight loss of the video quality.
—Again, since there is no detailed mechanism to adjust the size of compressedvideo data at the encoing process without modifying the default quality-asware mechanisms such as the quantization scale value, we maintain theAER to keep the transmission energy minimal to save the overall energy atthe source and at the destination at the least cost of the video quality.
5.2 Sensitivity of Error-Injection Rate and Adjusted Error Rate
(1) The effectiveness of EIR on the energy consumption and the video quality(Fig. 12)
(2) The effectiveness of AER on the energy consumption and the video quality(Fig. 12)
5.3 Energy/QoS tradeoff
(1) Extended tradeoff space for mobile video embedded systems
Energy Consumption and Quality
0
0.2
0.4
0.6
0.8
1
1.2
Enc EC Tx EC Enc EC +
Tx EC
Dec EC Rx EC Dec EC +
Rx EC
Quality in
PSNRN
orm
ali
ze
d R
ati
o o
f E
A-G
OP
to
GO
P
GOP-15 EAGOP-10
Fig. 10. Energy Reduction and Quality Degradation of EA-GOP compared to GOP(PLR = 0%, EIR = 10%, AER = 0% and -10%, FOREMAN 300 frames)
Transmission Energy Consumption with increasing EIR
0
0.2
0.4
0.6
0.8
1
1.2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
EIR (in %)
No
rma
lize
d E
ne
rgy
Co
ns
um
pti
on
EA-PBPAIR GOP-3 (baseline)
(a) Energy Consumption for TransmittingData
Video Quality with increasing EIR
0
0.2
0.4
0.6
0.8
1
1.2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
EIR (in %)
No
rma
lize
d Q
ua
lity
to
GO
P-8
EA-PBPAIR GOP-3 (baseline)
(b) Video Quality Evaluation
Encoding Energy Reduction with increasing EIR
0
0.2
0.4
0.6
0.8
1
1.2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
EIR (in %)
No
rma
lize
d E
ne
rgy
Co
ns
um
pti
on
EA-PBPAIR GOP-3 (baseline)
(c) Encoding Energy Consumption
Decoding Energy Reduction with increasing EIR
0
0.2
0.4
0.6
0.8
1
1.2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
EIR (in %)
No
rma
lize
d E
ne
rgy
Co
ns
um
pti
on
EA-PBPAIR GOP-3 (baseline)
(d) Decoding Energy Consumption
Fig. 11. Effects of Error Injection Rate on Energy Consumption and Video Quality(PLR = 10%, FOREMAN 300 frames, Each encoding is constrained with band-width)
(2) Effectively reduce the energy consumption of an encoding mobile device at thecost of the video quality
Video Quality vs. Energy Consumption at Source
0
5
10
15
20
25
30
15 20 25 30 35
PSNR (in dB)
En
erg
y C
on
su
mp
tio
n (
in J
ou
les
)
EA-PBPAIR PBPAIR
(a) Video Quality vs. Source Energy Con-sumption (Src EC)
Video Quality vs. Energy Consumption at Destination
0
5
10
15
20
25
30
15 20 25 30 35
PSNR (in dB)
En
erg
y C
on
su
mp
tio
n (
in J
ou
les
)
EA-PBPAIR PBPAIR
(b) Video Quality vs. Destination EnergyConsumption (Dst EC)
Fig. 12. Energy Consumption vs. Video Quality of EA-PBPAIR compared toPBPAIR (EIR = 1% to 50%, PLR = 5%, AER is set to Para1 = 15%, FOREMAN300 frames)
5.4 Feedback-based Error-Injection Control
(1) B. a set of experiments to show how our proposed feedback-based error-injectioncontrol works in order to keep the quality satisfied while minimizing the energyconsumption
(2) (B1) changing EIR can show how EIR works for adjusting quality of service fora fixed PLR—show how effectively our EA-PBPAIR can be used for adjusting quality by
adjusting EIR—base is a nominal EA-PBPAIR, i.e., a fixed EIR for error-injection
5.5 Adaptive EA-PBPAIR under Dynamic Network Status
(1) (B2) Also, adjusting EIR for varying PLR dynamically works well—Adapt EIR for EA-PBPAIR and Para1 for PBPAIR to satisfy the given
quality—base is a nominal EA-PBPAIR, i.e., a fixed EIR for error-injection
6. SUMMARY
REFERENCES
Video Quality vs. Source Energy Consumption
0
10
20
30
40
20 25 30 35 40
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(a) Video quality vs. Source Energy Consump-tion (AKIYO)
Video Quality vs. Destination Energy Consumption
0
2.5
5
7.5
10
20 25 30 35 40
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(b) Video quality vs. Destination Energy Con-sumption (AKIYO)
Video Quality vs. Source Energy Consumption
0
10
20
30
40
15 20 25 30 35
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(c) Video quality vs. Source Energy Consump-tion (FOREMAN)
Video Quality vs. Destination Energy Consumption
0
2.5
5
7.5
10
15 20 25 30 35
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(d) Video quality vs. Destination Energy Con-sumption (FOREMAN)
Video Quality vs. Source Energy Consumption
0
10
20
30
40
10 15 20 25 30
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(e) Video quality vs. Source Energy Consump-tion (COASTGUARD)
Video Quality vs. Destination Energy Consumption
0
2.5
5
7.5
10
10 15 20 25 30
PSNR (dB)
En
erg
y C
on
su
mp
tio
n (
Jo
ule
s)
EA-PBPAIR PBPAIR GOP-8
(f) Video quality vs. Destination Energy Con-sumption (COASTGUARD)
Fig. 13. Video Quality vs. Energy Consumption (Each encoding is constrainedwith bandwidth, EIR = 1% to 50%, PLR = 5%, AER is set to Para1 = 15%, Para2is varying, FOREMAN 300 frames)
Energy Saving at the Cost of Video Quality - akiyo
0
20
40
60
80
100
Source EC Destination EC
En
erg
y S
av
ing
(%
)
No Quality Loss
5% Quality Loss
10% Quality Loss
(a) AKIYO
Energy Saving at the Cost of Video Quality - foreman
0
20
40
60
80
100
Source EC Destination EC
En
erg
y S
av
ing
(%
)
No Quality Loss
5% Quality Loss
10% Quality Loss
(b) FOREMAN
Energy Saving at the Cost of Video Quality - coastguard
0
20
40
60
80
100
Source EC Destination EC
En
erg
y S
av
ing
(%
)
No Quality Loss
5% Quality Loss
10% Quality Loss
(c) COASTGUARD
Fig. 14. Energy Saving at the Cost of Video Quality by EA-PBPAIR
Effectiveness of Adaptive EE-PBPAIR on Video Quality
10
15
20
25
30
35
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Run Number
PS
NR
(in
dB
)
Adaptive EE-PBPAIR with varying EIR
Original EE-PBPAIR with fixed EIR = 30%
(a) Delivered Video Quality
Adaptive EIR
0
5
10
15
20
25
30
35
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Run Number
EIR
(in
%)
(b) Error Injection Rate
Fig. 15. Adaptive EA-PBPAIR with a decrease of EIR vs. EA-PBPAIR with afixed EIR
Video Quality
25
26
27
28
29
30
31
32
20 15 10 5 10 15
Packet Loss Rate (%)
PS
NR
(d
B)
Adaptive EA-PBPAIR Static EA-PBPAIR
(a) Delivered Video Quality
Adaptive EIR based on Feedback
0
5
10
15
20
25
30
35
20 15 10 5 10 15
Packet Loss Rate (%)
Err
or
Inje
cti
on
Ra
te (
%)
Adaptive EA-PBPAIR Static EA-PBPAIR
(b) Adaptive Error Injection Rate
Fig. 16. Adaptive EA-PBPAIR Robust to Varying PLR under Dynamic NetworkStatus