Transcript Slide 1

Human Factors and User Interfaces in
Energy Efficiency
Lin Zhong
ELEC518, Spring 2011
Motivation
User
User interface
Software
Application
Operating system
Hardware
Massive
Processor Memory
storage
Network
interface
Display &
other interface
hardware
2
Energy efficiency: definition
User productivity
Energy efficiency =
Avg. power consumption
= (User productivity) ×(Power efficiency)
Human-computer interaction
(HCI)
Low-power design
3
Limits
• Minimal power/energy requirements
• Human speeds
4
Speed mismatch
1000000
Times of improvement
Olympic Gold Metal winner: 100m dash (men)
100000
10000
Olympic Gold Metal winner: 100m dash (women)
# of transistors for Intel processor
Processor performance measured in MIPS
1000
100
10
1
1968
Sources: intel.com and
factmonster.com
1972
1976
1980
1984
1988
1992
1996
2000
2004
Year
A constantly slow user
An increasingly powerful computer
5
Slow-user problem
A computer spends most of its energy in
interfacing
1
Power (Watt)
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
Time (s)
Slow-user problem cannot be alleviated by
a “better” or more powerful interface
6
Model Human Processor
Three processes involved in the
user reaction to a computer
Perceptual process
Cognitive process
Motor process
Model Human Processor: Card, Moran & Newell’83
7
Perceptual process
• Fixations and saccades
– Fixation: information absorbed in
the fovea (60ms)
– Saccades: quick movements
between fixations (30ms)
– Each GUI object requires one
fixation and one saccade
• Rauding rate
– Raud: read with understanding
– 30 letters/second (Carver, 1990)
8
Cognitive process
• Hick-Hyman Law
– N distinct and equally possible choices
1
Cognitive delay  log 2 N  1 (s)
7
• Applicable only to simple cognitive tasks
– Selection: menu, buttons, list
9
General form
• Hick-Hyman Law
– pi : the probability that the ith choice is selected
1 N
1
Cognitivedelay   pi log (1 )
7 i 1
pi
– pi can be estimated based on history
10
Motor process
• Stylus operation
• Fitts’ Law
– A: distance to move
– W: target dimension along the moving direction
A
Motor delay  0.23  0.166 log 2 (  1) (s)
W
– Parameters adopted from (MacKenzie and Buxton, 1992)
11
Power Law of practice
• Speed on nth trial
– Sn = S1 na, where a ≈0.4
– Applies to perceptual & motor processes
– Does not apply to cognitive process or quality
50
45
40
35
30
Measurement
25
Power Law prediction
20
15
10
5
0
0
10
20
30
40
50
Learning curve of text entry using Twiddler, Lyons, 2004
12
Human capacity limitations
•
•
•
•
Perceptual
Cognitive
Motor
……
Human capacity
13
Cache
Speed mismatch
Cost to reduce
Task to outsource
Memory cache
Interface cache
CPU & memory
Computer & user
Memory access
latency
Frequently
accessed data
Interfacing energy
Frequent
interactions
Alleviate slow-user problem with a
“worse” or less powerful interface
14
Interface cache: examples
Flip phones
Average time spent on laptop per day
declined from 11.1 hours to 6.1 hours 5
months after Blackberry deployment
-----Goldman Sachs Mobile Device Usage
Study
15
Human thermal comfort
Starner & Maguire, 1999 and Kroemer et al, 1994
16
A hot case: 3-Watt Nokia 3120
Every One Watt increases surface temperature
by about 13 deg C
Phone case temperature will be
40 deg C higher.
17
Minimal power/energy requirement
Visual and auditory output
Emin ≈ Ω·D2·10-13 (Joule)
D
Point source
Ω
About 10-14 (Joule) for most
handheld usage
Minimal energy requirement for
1-bit change
with irreversible computing
10-21 (Joule)
(Landauer, 1961)
18
Insights for power reduction
P∝
D
Point source
Ω
Ω·D2
η(λ)·V(λ)
λ: wavelength of light/sound
η(λ): conversion efficiency
from electrical power
V(λ): relative human
sensitivity factor
Reflective layer
to control Ω
19
Text entry speed (productivity)
180
Speed (words per minute)
160
150
140
Raw speed
120
Corrected speed
100
80
60
40
25
23
22
13
20
12
15
7
0
Speaking
mini hardware keyboard
Software keyboard with
stylus
Handwriting
20
Impact of human factors
1
Power (Watt)
0.8
Using Calculator on
Sharp Zaurus PDA
0.6
0.4
0.2
0
0
1
2
3
4
5
Time (s)
Length of idle periods cannot be significantly reduced
Power consumption in idle periods is dominated by interfacing devices
99% time and 95% energy spent in idle periods during interaction
21
Experimental setup
Devices
HP iPAQ 4350
Sharp Zaurus SL-5600
Windows
Transflective/back light
Bluetooth
Speech recog.
Linux/Qt
Reflective/front light
Intel Xscale 400Mhz
240X320, 16-bit color
mic., speaker & headphone jack
22
Experimental setup (Contd.)
Measurement
Host machine
GPIB card
GPIB cable
Agilent 34401A
multimeter
Vs Rs
iPAQ H3870
Vdd
5V
200 samples/second
23
Experimental setup (Contd.)
1.6
1.6
1.2
1.2
Power (W)
Power (W)
Write “x” with
stylus/touchscreen
0.8
0.4
0
0.8
0.4
0
0
0.5
1
1.5
0
Time (s)
0.5
1
1.5
Time (s)
Extra energy consumption by
writing “x”
Extra energy/power consumption of an event is
obtained through differential measurement
24
Power breakdown
Power consumption (mW)
4
A handheld usually spends
most time being idle but the
display has to be on most
time
Earphone
3
Speaker
Lighting
LCD
2
If the display is not on, the
speaker subsystem is usually
on
Computing
Basic idle
1
0
iPAQ
Zaurus
Computing: carrying out DCT repetitively
25
Energy characterization
• Visual interfaces
– Graphical user interfaces (GUIs)
– Digital camera
• Auditory interfaces
– Recording/playback
– Speech recognition & synthesis
• Manual text entry
26
GUIs
• Stylus/Touch-screen
• Most energy/time spent in idle periods
– Energy consumed by computing negligible
• Task time determines energy consumption
1
Power (Watt)
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
Time (s)
27
Speech synthesis & recognition
• Infer the behavior of Voice Command by
comparing voice recording and power
trace
• Computing is not demanding
• Used as baseline for comparison
Voice recording
2
Power (W)
1.6
1.2
0.8
Power trace
0.4
0
1
207
413
619
825
1031
1237
Time (1/206 s)
1443
1649
1855
2061
2267
28
Comparison: Output
• Speech is better only when
iPAQ
2
– display is turned off
– earphone is used
– nighttime usage
Lighting required for text
routput
Lighting not required for text
1
Energy efficiency
ratio
Rspk Ptxt
r

Rrd Pspk
0
display off
earphone
display on
earphone
display off
loudspeaker
display on
loudspeaker
Different scenarios
If r >1, speech output is more energy-efficient
29
Comparison: Text entry
100
HW MKB-ideal
VKB-ideal
Letter Recog.-ideal
HW MKB
VKB
Letter
State of
the art
rinput
10
Near
future
Ideal
1
Speech recog. input rate (cwpm)
0.1
0
20
40
60
80
100
If r >1, speech recognition is more energy-efficient
120
140
160
30
Comparison: Text entry (Contd.)
100
HW MKB-No LCD
VKB-No LCD
Letter Recog.-No LCD
HW MKB-No LCD/Night
VKB-No LCD/Night
Letter Recog.-No LCD/Night
rinput
10
1
Speech recog. input rate (cwpm)
0.1
0
20
40
60
80
100
120
140
160
Handwriting recognition is inferior to alternatives
Speech recognition can be the most energy-efficient
31
Comparison: Command & control
Maximal no. of words per command
• Speech vs. GUI operation
9
8
Ideal
7
95% accurate
95% accurate/No LCD
6
Assume each stylus
tapping takes 750ms
95% accurate/No LCD/Light
5
4
3
2
1
0
1
2
3
4
5
No. of taps
Single word voice command is more energy-efficient than GUI
operation with 2 taps
32
Observations
• User productivity (speed) is critical
– energy consumed being idle is significant
• Handwriting-based text entry is inferior
• Speech-based text entry can be superior
– Turning off display is important
– Accuracy
• Loudspeaker consumes significant power
– Earphone incurs usability issue
– Wireless audio delivery not energy-efficient
• “Computing” usually consumes trivial energy
33
Examples of energy inefficient interfaces
Kyocera KX2325
LG VX 6100
Microsoft Voice
Command 1.01
34
Energy efficiency: definition
User productivity
Energy efficiency =
Avg. power consumption
= (User productivity) ×(Power efficiency)
Human-computer interaction
(HCI)
Low-power design
35
Model of Man
• Herbert Simon
– Turing Award (1975)
– Nobel Prize in Economics (1978)
• Human mind is simple; its apparent complexity is
due to the environment’s complexity
– Short-term memory is fast but small (~7)
– Long-term memory is unlimited but writing takes time
(10 to 30 seconds)
– Retrieval from long-term memory is associative and
depends on the storage structure
Bounded rationality
• Limitation on ability to plan long behavior
sequences
• Tendency to set aspiration levels for each goal
• Tendency to operate on goals sequentially
rather than simultaneously
• Satisficing rather than optimizing search
behavior
http://www.princeton.edu/~smeunier/JonesBounded1.pdf