Post on 03-Nov-2021
transcript
Shuo Deng, Hari Balakrishnan
MIT CSAIL
1
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
News IM Micro-blog Game Email SocialNetwork
Finance
En
erg
y P
erc
en
tag
e
Data Active, No Data High PowerIdle State Switch
Problem: 3G/LTE is a battery hog “Up to 14 hours on 2G” “Up to 6.5 hours on 3G”
2
Goal: Reduce Energy Consumption
Context: Radio Resource Control
3
Cell_DCH (Active)
Cell_FACH (High Power Idle)
Cell_PCH/IDLE (Idle)
Inactivity Timer t1
Inactivity Timer t2
Power Consumption
4
IDLE
Active
High Power Idle
IDLE
t1
t2
t1
t2
IDLE
Fast Dormancy
5
Active
Idle
Inactivity Timer t1
Inactivity Timer t2
High Power Idle
Active Idle
Challenges Switching between states takes time(1~3 seconds), and
consumes energy
Signaling overhead
6
IDLE
t1
IDLE t2
Active Idle
Contributions A traffic-aware design to control radio state transitions
to reduce energy consumption
MakeIdle: when to switch to Idle
MakeActive: when to switch to Active
Experimental evaluation on real usage data
Energy reduction up to 75% across different carriers 7
Active Idle
MakeIdle
MakeActive
System Design
8
Control Module
App1 App2 App3
Modified Socket Layer
Fast Dormancy Interface
3G Radio
Socket Calls
Packet/Socket Call Information
Trigger Fast Dormancy
Data
Switch to Idle
Switch to Active
Idle Active
9
Power
time
IDLE
t1
IDLE t2
MakeIdle Algorithm
10
time
Power
> If , should switch to Idle mode to minimize the energy consumption.
IAT time
Power
If IAT > threshold, should switch to Idle mode.
Idle Active
MakeIdle
Inter Arrival Time
MakeIdle Algorithm Predict whether the IAT will be greater than threshold
Wait for a short period of time twait, if no packet comes, then put the radio to Idle mode
Why: the longer the network is idle, the longer it is likely to remain idle
11
Idle Active
MakeIdle
MakeIdle: Picking twait
12
…
Previous N packets
)( ___ switchstatenoswitchstate EnergyEnergy
Idle Active
MakeIdle
P(IAT>t+threshold | IAT>t)
13
time
Power
time
Power
time
Power
Current Scheme
With MakeIdle
With MakeActive
MakeActive Algorithm Reduce the number of state switches by introducing a
small delay when the radio is in Idle mode and data transmission requests come from the mobile device side
How much delay for each request?
Fixed delay bound
Learning algorithm
14
Idle Active
MakeActive
Evaluation Setup Energy profiling
Power consumption profiles for 4 US major carriers: AT&T, Verizon, T-Mobile, Sprint
Trace driven simulation
Tcpdump traces for real usage data, collected from 9 users, 28 days in total
15
1 2 3
Evaluation: MakeIdle
16
-10
0
10
20
30
40
50
60
70
80
1 2 3
En
erg
y S
aved
(%
)
User ID
Incr
ease
d S
ign
alin
g
Ove
rhea
d
User ID
40x
35x
30x
25x
20x
15x
10x
5x
0
95% IAT across users
95% IAT per user
MakeIdle
Prior knowledge of IAT (Oracle)
0
5
10
15
20
25
30
1 2 3
Evaluation: MakeIdle
17
Prediction Accuracy
Idle Active
MakeIdle
Fals
e P
osi
tive
(F
P)
or
Fals
e N
egat
ive
(FN
)
95% IAT across users FP
95% IAT per user FN
MakeIdle FP
95% IAT across users FN
95% IAT per user FP
MakeIdle FN
User ID
Evaluation: Different Carriers
18
-10
0
10
20
30
40
50
60
70
80
T-Mobile AT&T Verizon 3G Verizon LTE
En
erg
y S
aved
(%
)
Carriers
95% IAT across users
95% IAT per user
MakeIdle
MakeIdle+MakeActive
Related Work Inactivity timer reconfiguration
Statistical method [Falaki et al, 2010]: 95 percentile packet inter arrival time
Applications-Involved Design
TailEnder [Balasubramanian et al, 2009]: each application specifies its delay tolerance
TOP [Qian et al, 2010]: application predict the gap between its own traffic transmissions
TailTheft [Liu et al, 2011]: application specifies delay tolerance and predicts transmission duration
19
Conclusion A traffic-aware design to control state transitions of
3G/LTE radio to reduce energy consumption on mobile devices
Require no modifications of the applications
Save 3G/LTE energy consumption by up to 75% across different carriers
20