Transcript slides

Energy and Performance of
Smartphone Radio Bundling in
Outdoor Environments
Ana Nika*, Yibo Zhu*, Ning Ding+, Abhilash Jindal+, Y. Charlie Hu+
Xia Zhou^, Ben Y. Zhao* and Haitao Zheng*
*UC Santa Barbara, ^Dartmouth College, +Purdue University
[email protected]
Wireless Bandwidth Crisis
• Many data-hungry applications for smartphones
How do we support these new applications?
1
Single Radio Not Enough
Cellular (LTE)
• Optimal: 75Mbps
• In practice:
outdoors 2%-20%
o movement
o congestion
2
Single Radio Not Enough
Indoors
WiFi
60GHz
WiFi: ~150Mbps
60GHz: ~6Gbps
What about outdoors?
WiFi suffers from interference, poor quality
60GHz small range, easily blocked
3
Radio Bundling
LTE
WiFi
1Mbps
2Mbps
Radio
Bundling:
3Mbps
4
Our Goals
1. What is the maximum benefit that can be
provided by bundling on today’s devices?
1. What is the energy cost of bundling?
Via real measurements using a smartphone app
in 5 US cities and 63 outdoor locations.
5
Existing Work
• MPTCP: transport-layer bundling implementation
– non-optimal (fair to non-MPTCP users)
– not designed for wireless radios
– does not consider energy
• Our goal is to study optimal radio bundling
– understand fundamental limitations of bundling
– how different radio usage options compare
– how we can improve current implementations
Energy
Bundling
LTE-only
?
MPTCP ?
Bundling
WiFi-only
?
Throughput
6
Methodology
• Android-based app
– turn on WiFi & LTE radios simultaneously
Not available
by default
• Power model for bundling and non-bundling
– take into account CPU & WiFi, LTE radios
• Measurements at 5 cities & 63 outdoor locations
– file download/upload (0.5MB-5MB in size)
– derive throughput and energy of different radio usage options
– capture different RF conditions
o
in terms of
THROUGHPUTWiFi
THROUGHPUTLTE
7
Radio Usage Options
• Bundling
– Optimal traffic partitioning
o
trace playback analysis
– MPTCP
• Non-bundling
– Single Radio: LTE-only, WiFi-only
– Best Radio: max (LTE-only, WiFi-only)
– Radio Switching: instantaneous switching between two
radios
8
Energy Profiling
• Accurate signal-strength aware power model
– componentized: WiFi, LTE radios & CPU
– takes into account energy tail
– < 8% error rate
• Power draw projection over time
Power (W)
Bundling
3.5
3
2.5
2
1.5
1
0.5
0
CPU
WiFi
LTE
0
2
Energy tail
Data Transfer Ends
4
6
8
Time (s)
10
12
14
9
Outline
• Motivation
• Methodology
• Throughput Performance
• Energy Consumption
• Conclusion
10
Radio Frequency Conditions
THROUGHPUTWiFi
q=
THROUGHPUTLTE
q
q =10
-2

LTE 100x better than WiFi
q =10
2

WiFi 100x better than LTE
11
Ratio over Ideal Bundling
Ratio over Ideal Bundling
Bundling’s Throughput Improvement
Best Radio
Bundling
Switching
Bundling
1
Gain:
2x – 5x
0.8
0.6
0.4
0.2
0
10-2
1
10-1
100
101
Relative WiFi/LTE Throughput Ratio
102
MPTCP
Bundling
MPTCP:40
%-85% of
optimal
bundling
0.8
0.6
0.4
0.2
0
10-2
10-1
100
101
Relative WiFi/LTE Throughput Ratio
102
12
Sources of Bundling Gain
• Radio independence
– LTE and WiFi radios do not interfere with each other
– dual-core CPU supports both radios
• Traffic partitioning matters!!!
– accurately projects radio throughput to fully utilize both radios
100
MPTCP performs
poorly.
CDF (%)
80
60
An approach with
high noise
performs better
than MPTCP.
40
20
0
0
0.2
0.4
0.6
0.8
Ratio over Ideal Bundling
1
13
Energy Analysis
• LTE is the heaviest energy drainer
– consumes at least 50% of the total power draw
• CPU consumes a significant amount of energy
– 60% for WiFi-only
– 20%-23% for Bundling, MPTCP, LTE-only
3.5
3
2.5
2
1.5
1
0.5
0
WiFi-only
Power (W)
Power (W)
Bundling
CPU
WiFi
LTE
0
2
Data Transfer Ends
4
6
8
Time (s)
10
12
14
3.5
3
2.5
2
1.5
1
0.5
0
Data Transfer Ends
CPU
WiFi
0
1
2
3
4
5
6
7
Time (s)
14
Bundling Energy Cost
Bundling Energy Cost
Bundling Energy Cost
• Energy cost of a file transfer
4
Bundling
Best Radio
3
2
•
Bundling energy cost ≤
LTE-only cost
•
Bundling energy cost
> WiFi-only cost,
because of LTE radio
•
MPTCP sometimes
saves energy
LTE-only
WiFi-only
1
0
10-2
4
3
10-1
100
101
Relative WiFi/LTE Throughput Ratio
102
Bundling
MPTCP
o Bundling consumes
more CPU energy
2
1
0
10-2
10-1
100
101
Relative WiFi/LTE Throughput Ratio
102
15
Bunding Energy Cost
Energy Cost vs. Throughput Gain
4
WiFi-only
LTE-only
MPTCP
3
2
Potential bundling
implementation
1
0
1
10
100
Bundling Throughput Gain
•
•
Bundling improves transfer throughput & transfer ends earlier
Energy saving due to reduced transmission time
– compensates the extra power draw of the additional radio
16
Conclusion
• Bundling is highly beneficial, achieve 2x-5x improvement over
single radio
• MPTCP achieves only a portion of the total performance
possible
• Bundling has higher instantaneous power draw, but can lead to
lower energy cost due to faster transmission
• Our accurate componentized power model identifies the
significant role CPUs play in energy usage
• There is ample room for a new bundling protocol that provides a
better tradeoff between performance and energy usage
17
Thank You!
Questions?
18
Impact of File Size on Throughput
100
0.5MB
1MB
2MB
4MB
5MB
CDF
80
60
40
Switching
20
Best Radio
0
0
0.2
0.4
0.6
0.8
1
Ratio of Online Selection
over Ideal Bundling
100
0.5MB
1MB
2MB
4MB
5MB
CDF
80
60
40
• Consistent across
all transfer sizes
• Slightly higher for
512KB and 1MB
transfers
o TCP slow start
20
0
0
0.2
0.4
0.6
0.8
1
Ratio of MPTCP over Ideal Bundling
19
Impact of File Size on Energy
100
CDF
80
60
0.5MB
1MB
2MB
4MB
5MB
40
20
0
0
1
2
3
4
5
Bundling Cost over Best Radio
100
• Consistent across
all transfer sizes
• Higher for 512KB
and 1MB transfers
o LTE energy tail
CDF
80
60
0.5MB
1MB
2MB
4MB
5MB
40
20
0
0
1
2
3
4
5
Bundling Cost over MPTCP
20