T-110.6120 Energy-efficient Mobile
Computing
Energy-efficient wireless
networking
23.4.2014
Matti Siekkinen
Outline
•
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
Conclusions
Energy efficiency of wireless and mobile
networking
•  In short: Resulting battery life when using smartphone to
access the Internet
•  Two concepts
–  Wireless communication: use radio(s) to communicate
–  Mobile networking: move while communicating
•  Using radio requires a certain amount of power
–  How much depends on the type of wireless technology
•  Basically the PHY and MAC layers
•  Being mobile means that this power is not constant
–  We’ll come to the causes and consequences later on…
Questions, questions…
•  Q: Which one is more energy efficient?
–
–
–
–
3G, WLAN, or LTE?
Lumia 920 or iPhone 5?
P2P or C/S?
...
•  A: It depends...
4
A glance at power consumption…
3.000
(5,2.249)
2.500
(211,2.516)
?
?
1.500
(10,1.281)
?
(113,1.494)
1.000
0.500
WLAN
WCDMA
0.000
1
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100
105
110
115
120
125
130
135
140
145
150
155
160
165
170
175
180
185
190
195
200
205
210
215
220
225
230
235
240
243
245
249
Power(Watt)
2.000
Time(second)
Watching YouTube from N95
5
Energy efficiency of wireless and mobile
networking
•  Energy efficiency: Spend as few Joules over a period of
time as possible
–  Minimal average power consumption
•  How to improve the efficiency by means of software?
–  Switch off unnecessary hardware
•  Some of the radio circuitry
–  Increase the number of bits transmitted/received per Joule spent
–  Reduce the number of bits to transmit/receive
Outline
•  Intro
•  Standard power management techniques
–  Wi-Fi
–  Cellular networks
–  Tail energy
•  Beyond standard techniques
•  Example optimizations
•  Conclusions
Standard power management techniques
•  Operation defined in a standard document
•  Usually
–  implemented by each and every device
–  requires cooperation between mobile device and the network
•  Wi-Fi: IEEE 802.11 standard
•  Cellular networks: 3GPP releases
–  UMTS (3G, Rel-99)
–  LTE (4G, Rel-8)
WNI states and transitions
•  Management of wireless network interface happens through
different states
–  Set of states are technology specific
•  WNI transitions from state to another according to some rules
–  Promotions based on traffic demand
–  Demotions usually timer specified
•  What states?
–  E.g. receive, idle, and sleep in WiFi
–  Correspond to specific modes of the hardware
–  CELL_DCH, CELL_PCH etc. for 3g
•  Correspond to different kind of resource allocation (i.e. channel type)
•  States have different power characteristics
–  Part of circuitry can be powered off at run time (sleep)
CHAPTER 2. LTE
13
For real-time streaming applications such as voice calls, the data sent
during every transmission and the bandwidth utilized are very small. For
applications transmitting small amount of data for many times semi persistent scheduling is more adaptable in which the UE does not need to request
for a Grant each time for transmission of data. Instead, the UE is provided
Traffic
RRC Inactivity
with a transmission pattern which it follows and transmits the data during
burst
timer
that particular time slot. For example, during a voice
call running
the UE sends an
initial SR and gets the transmission pattern from the scheduler after which
the data is sent on the allocated time. This reduces the complexities and
overhead of the scheduler and the UE.
There are many di↵erent types of scheduling algorithms which are designed based on the need of vendors. Commonly used algorithms include
Round Robin [42], Proportional Fair [47], Maximum Channel Quality Index
(CQI) [43], Channel aware scheduling algorithm
[42].
RRC_IDLE
Wi-Fi, 3G, LTE: different power states
2.3
LTE UE and RRC States
In LTE, a UE can shift between three states namely, RRC DISCONNECTED,
RRC CONNECTED and RRC IDLE.
SLEEP
RECEIVE
PSM TRANSMIT
Timeout
IDLE
Figure 2.4: States of LTE UE
Initially, when the UE is in the OFF state, it does not hold any active connection with the eNodeB. The state at which the UE is turned on
and is searching for a possible base station for registration or the phase
when the UE is in Airplane mode could be termed as RRC DETACHED
Outline
•  Intro
•  Standard power management techniques
–  Wi-Fi
–  Cellular networks
–  Tail energy
•  Beyond standard techniques
•  Example optimizations
Wi-Fi (802.11) power consumption
•  Power consumption depends on operating mode
Energy = Power(operating mode)* Duration(operating mode)
Continuously Active Mode (CAM)
Power Saving Mode(PSM)
SLEEP PS
IDLE
RECEIVE
PR
RECEIVE
TRANSMIT
TRANSMIT
PT
IDLE PI
12
Wi-Fi operating modes and power
•  Order of magnitude less power drawn in sleep state
WNI operating mode
Average Power (W)
Nokia N810
HTC G1
Nokia N95
IDLE
0.884
0.650
1.038
SLEEP
0.042
0.068
0.088
TRANSMIT
1.258
1.097
1.687
RECEIVE
1.181
0.900
1.585
Wi-Fi power saving
•  Allows (part of) Rx/Tx circuitry to be temporarily shut
down
•  Coordinated with the AP
1.
2.
3.
4.
Node-to-AP: “I am going to sleep until next beacon frame”
AP knows not to transmit frames to this node, buffers them
Node wakes up before next beacon frame
Beacon frame: contains list of mobiles with AP-to-mobile
frames waiting to be sent
•  Traffic Indication Map (TIM)
5.  Any frames buffered for the node?
•  Yes à request for them from AP and stay awake until received
•  No à go back to sleep until next beacon frame
14
Wi-Fi power saving (cont.)
•  Standard PSM is poison for interactive applications
–  Frequent transitions to and from sleep mode adds lot of delay
•  “Adaptive” version used in practice
–  Use timer: if no frames for 100-200ms, then sleep
–  Timer value is device specific
PSM timeout
Rx
Tx
PSM timeout
Idle Sleep
15
Outline
•  Intro
•  Standard power management techniques
–  Wi-Fi
–  Cellular networks
–  Tail energy
•  Beyond standard techniques
•  Example optimizations
•  Conclusions
CELL-­‐ PCH T2 pagingCEL
T3 R
Fast Dormancy IDLE –  Consequently, power consumption of a
mobile phone too
T1 CELL-­‐ FACH •  Radio resources (channel usage)
controlled by the Radio Resource
Control (RRC) protocol
bulk low vol.
data traffic
CELL-­‐ DCH 3G power management
Timer expiry Rx/Tx Saving Energy in Mobile Devices Through TCP Receive B
•  Four states and three inactivity timers
–  States correspond to transport channels
•  Dedicated channel (DCH),T1Forward access
channel (FACH), Paging channel (PCH)
–  In practice, timers
are not just single
CELL_DCH
CELL_FACH
parameter
Data Activity
Da
t
a
T2
Ac
tiv
ity
•  e.g. track average
nb of bytes over time
Da
ta
window and useActhresholds
CELL_PCH
tiv
ity
•  typically at least a few seconds long
low power high power CELL_DCH
CELL_FACH
T3,
–  Operator controlled,
phone cannot
change
RRC Connect
RRC Disconnect
CELL_PCH
IDLE
RRC with Four States
Fig. 1: 3G state transitions and power consumption in di↵erent states.
LTE power management
•  Same concept than in 3G
•  Simplified RRC protocol
–  Only two states: RRC_CONNECTED and RRC_IDLE
–  Inactivity timer to switch to RRC_IDLE
•  typical value: 10s
Traffic
burst
RRC Inactivity
timer running
RRC_CONNECTED
current drawn by
LTE smartphone
RRC_IDLE
Outline
•  Intro
•  Standard power management techniques
–  Wi-Fi
–  Cellular networks
–  Tail energy
•  Beyond standard techniques
•  Example optimizations
•  Conclusions
Tail energy
•  All wireless network interfaces exhibit tail energy
–  Energy spent being idle with radio on à wasted energy
•  Due to inactivity timers
–  Mandate how long radio remains in active state (rx on) before state
transition to inactive state (rx (partly) off)
•  Timers are necessary
–  Sporadic communication patterns might lead to very frequent transitions
–  State changes require signaling between phone and base station
•  Transmitted on shared channel with limited capacity
•  Signaling traffic volumes must be limited
–  Also switching between hardware modes adds some delay
•  Timer values vary between technology
–  Wifi≈100-200ms
–  3G and LTE: in the order of seconds (varies between ISPs)
How to minimize tail energy?
•  Wi-Fi tail is already short
–  No need for specific mechanisms
•  Fast Dormancy for 3G
–  Cuts tail duration down to 3-5s
•  DRX/DTX for LTE
–  Especially cDRX/cDTX: connected mode discontinuous reception/
transmission
–  Typically cuts tail duration down to a few hundred milliseconds
•  3G has also CPC
–  Continuous packet connectivity
–  Similar to LTE’s cDRX/DTX
–  Introduced in Rel-7 but often not (yet?) fully supported by deployed
networks and devices
Fast Dormancy (3G)
•  Two flavors: legacy and standard
•  Legacy came first and is phasing out
–  Phone transmits SIGNALING CONNECTION RELEASE
INDICATION msg à tears down PS signaling connection
•  Normally phone uses to communicate some error conditions
–  Good: Results in immediate transition to low power IDLE
–  Bad: new communication requires re-establishing of signaling
connection à frequent use causes signaling storms
•  Standard FD in Rel-8 is network controlled
–  Phone requests network to transition it into an appropriate state
(e.g. CELL_PCH)
–  Network either allows or denies and decides appropriate state
•  E.g. too frequent requests are rejected
not receive any data for a certain time, it enters the DRX cycle
phase and activates the Short DRX cycle with Short DRX
timer (Ts). In the Short DRX cycle state, the UE monitors the
PDCCH during the DRX OnTimer period.
Connected mode DRX/DTX (LTE)
1. States of LTE UE
•  DRX works in LTE’s Fig.connected
state
–  Hence, also called
cDRX (connected
Initially, when the UE is in the OFF state, it does not hold
mode DRX)any active connection with the eNodeB. The state at which the
UE is turned on and is searching for a possible base station
for registration or the phase when the UE is in Airplane mode
could be termed as RRC DISCONNECTED state. Once the
–  Check periodically
new
data
is waiting
UE finds base if
station
coverage,
it registers
to the Mobility
Entity (MME) through the LTE Attach proce–  Very similarManagement
to
PSM
in 802.11
dure. After the registration is successful, the UE moves to the
RRC CONNECTED state. In the RRC CONNECTED state the
RRC inactivity timer (T )
arriving data UE has a Radio Resource Control (RRC) connection with theidle
eNodeB and maintains an active connection through various
Fig. 2. RRC State transition of LTE UE
C
ONNECTED
STATE
IDLE STATE
signaling messages. Once the UE sleeps for a longer duration
on
without any data transmission, it moves to the RRC IDLE
The Short DRX timer is active for a time period called as
state after the expiry of inactivity timer (Inactivity). When
the UE is out of coverage area, or during the process of the Short DRX timer period (TS)after which the UE activates
off
area update it de-registers with the eNodeB and moves to the Long DRX timer (Tl). In Long DRX cycle, the wakeup
duration of the UE
for monitoring the PDCCH is less frequent
DRXRRC
inactivity
transition
DISCONNECTED state.
DRX cycle DRX on-duration
cycle. Once the UE enters the Long DRX
is the
that was than the Short DRXto
timer DRX
(DRX
) saving mechanism in LTE, timer
idlepower
(DRX shifts
) from theIdle
RRC CONNECTED to RRC IDLE
(DRX
c) consumption by cycle, it on
introduced in order to reduce
the power
making the UE to move to sleep and idle states when there mode if the Inactivity timer expires. When the UE gets an
is no data transmission to and from the UE based on specific intimation about data to be received through the PDCCH, or
when it needs to transmit data, it moves from RRC IDLE
timers.
In order to obtain information about scheduling, the User mode to RRC CONNECTED mode. Once all the data is
Equipment (UE) monitors the physical down-link control transmitted, the UE again activates the DRX cycle in the
•  DRX operates in cycles
Rx
Rx
Outline
•
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
Conclusions
Is there room for optimization?
•  Typical application consumes a lot more energy than is
strictly necessary
–  Even with standard power saving mechanisms
•  Three reasons:
1.  Radio hardware is not perfectly power proportional to the
offered load
2.  Energy utility of wireless communication is context dependent
3.  The underlying hardware power management mechanisms are
rarely optimal for the applications being used
25
Power (dis)proportionality
•  Power draw does not scale linearly with amount of work
done
–  Bits transmitted/received per Joules spent typically increases
with data rate
•  Idle power consumed by hardware just being powered
on
–  That constant power added regardless of transmission rate
•  Idle power takes a larger share of slower transmission
–  Over-the-air (OTA) data rate of wireless channel != throughput
–  Each packet may be transmitted continuously at OTA rate but
idle time in between packets spent as tail energy
Energy efficiency is context dependent
Context
Type of WNI
used
energy utility of
data transmission
energy spent by
connectivity mgmt
Quality of wireless
network access
Network load
Signal quality
throughput
•  Moving user experiences varying signal quality
data rate
(modulation)
tx power
–  Dynamic switching of modulation based on SNR à data rate changes
–  Poor SNR à link layer retransmissions
–  Worse SNR requires more transmit power
•  Affects also reception (cont. tx of signaling msgs)
•  Static user may experience varying network load
–  Other users take up air time
Limits of underlying hardware power
management mechanisms
•  Hardware power management mechanisms
implemented at PHY/MAC layers
–  The protocol stack layers that interface the hardware
•  Separation of concerns in the layered protocol design
–  Layer is concerned only about its own responsibilities and
functionalities
–  Cross-layer protocols are not so common
•  At least not those that cross the stack up to application layer
•  à power management mechanisms are completely
unaware of application behavior
–  Same mechanism regardless of the type of application
How can we do better?
•  Traffic scheduling
–
–
–
–
Shape traffic to improve energy utility (bits per joules)
Reduce energy spent due to contention
Take context dependency into account
Handling background traffic energy efficiently
•  Smart use of wireless network interfaces
–  Smartphones have many
–  Use always the most energy efficient one
–  Often requires ability to predict connectivity
•  Application specific optimizations
Outline
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
–  Traffic shaping in multimedia streaming
–  Optimized prefetching of video streaming
–  Context-aware scheduling of transfers
•  Conclusions
(a) On N95 and N810(16-32KBps).
(b) On N95 a
EnergyFig.
utility
6. Energy utility of TCP upload with single flow a
2500
Measurements
2000
using Samsung
Nexus
1500 S
Wi-Fi download (PSM)
Energy Utility(KB/J)
Energy Utility(KB/J)
2500
2000
1500
1000
500
Measured
Estimated
512
1024
1536
2048 2854(no-trickle)
Data rate(KBps)
1000
500
512
102
(a) TCP download.
(b
•  Energy utility improves with data rate
à should
transmit
as fast
possible
Fig. always
7. Energy
utility
of as
TCP
download/upload on Ne
TCP flows. The biggest value on the X-axis is the m
point shows the mean and the standard deviation of
TABLE 9
Energy consumed by mobile streaming
•  Mobile media streaming drains battery quickly
–  Constant bit rate multimedia traffic is not energy friendly
–  Forces the network interface to be active all the time
Mobile Internet Radio
power draw on E-71
(TCP-based streaming)
Datarate
(kBps)
Start-up
Time (s)
8
WLAN power (W)
3G power (W)
PSM
CAM
48kBps
2Mbps
18
0.53
1.06
1.30
1.30
16
10
0.99
1.07
1.30
1.30
24
10
1.04
1.07
1.27
1.35
•  Idea: Shape traffic into bursts so that it is more energy efficient
to receive
–  Bursts sent at high data rate
–  “Race to sleep”
Multimedia Traffic Shaping with Proxy
•  Client sends request to proxy
•  Proxy
Fast Start phase to quickly
fill up playback buffer
–  forwards request to radio server
–  receives and buffers media stream
–  repeatedly sends in a single burst to client
•  802.11
–  PSM is enabled
–  WNI wakes up to receive a burst at a time
–  Waste only one timeout per burst
•  3G & LTE
–  Long enough burst interval
à inactivity timers expire
à switch to lower power state or activate DRX in
between bursts
33
Where is the free lunch?
•  Typical multimedia traffic leaves lots of network capacity
unused
•  Several vantage points possible for traffic shaper proxy
placement
wireless
wired
HSPA/LTE
8-3
Wi-Fi
bps
Internet
> 100 Mbp
50 M
video
server
s
Web
server
•  Intuition: Use as large as burst size
as possible
–  Maximize sleep time in between bursts
Bytes
What is the right burst size?
Burst,
BS=rsxT
= buffer
Burst
Idle period
between two
consecutive
Serverbursts, tidle
Duration,
tbd
Phone
streaming Tim
client
application
Burst Interval,
streaming
server T
application
1000not fit entirely à
Burst, burst will
–  Too large
sxT
800
TCP BS=r
flow
control kicks
in
600
Idle period
–  BufferBurst
is drained at
stream encoding rate
between two
Duration,
400
à
Excess
content
will be received at
consecutive
tbd
200
bursts, tidlethat rate
Burst Interval, T
3s
192 kbps
124 kbps
0
•  Lower energy
Time utility 0
1
2
3 4 5 6 7 8
Burst Interval, T (s)
9 10
200
KBytes
Avg. Power Consumption (mW)
Saving EnergyTCP
in Mobilereceive
Devices Through
TCP Receive
Multimedia Streaming
•  Problem:
buffer
is ofBu↵er AwareTCP
limited size
1200
150
zz
zz
2
100
50
1
•
1:5
TCP
zz
3
4
too large burst
! TCP flow
ctrl active
4:17 4:18 4:19 4:20 4:21 4:22
Time (S)
g. 5: Bursty traffic pattern and the idle Fig. 6: Power consumption of Nokia E-71 Fig. 7: The burst received by E-71 when
riod between two consecutive bursts.
at di↵erent burst intervals via Wi-Fi.
T = 10 s.
being present with each packet in slow constant bitrate streaming. The bar chart in Figure 6 shows that
What is the right burst size?
Burst,
Bytes
•  Burst size that offer maximalBS=r
energy
sxT
savings exists
–  Option 1: Make burst
size match
Idle period
Burstexactly
receive buffer size
between two
Duration,
•  Max burst sizeconsecutive
= playback buffer
tbd size
+TCP receive bursts,
buffert size
idle
–  Option 2: Max burst interval & size limited
by amount of initially buffered content Time
Burst Interval, T
•  Cannot let the playback buffer run dry
5: Bursty
traffic
pattern
and the to
idle
•  Optimality ofFig.
such
burst
size
is easy
period between two consecutive bursts.
prove
Avg. Power Consumption (mW)
Saving Energy in Mobile Devices Through TCP Receive Bu↵er Aw
1200
192 kbps
124 kbps
1000
800
600
400
200
0
0
1
2
3 4 5 6 7 8
Burst Interval, T (s)
9 10
Fig. 6: Power consumption of Nokia E-71
at di↵erent burst intervals via Wi-Fi.
–  Check: M. Hoque et al. Saving Energy in
to being
with each packet in slow constant bitrate streaming. T
Mobile Devices
forpresent
On-Demand
consumption
of Nokia E-71 decreases as the burst interval T or
Multimediapower
Streaming
- A cross-layer
duration
of the Wi-Fi
Approach. the
ACM
TOMCCAP.
2014.inactivity timer is 200 ms and thus tail ener
tail energy phenomenon can hurt much more if the idle period between
the size of the bursts are small, especially when streaming via cellular ne
timer values are 3many
seconds. These inactivity timers would never exp
s
and
zz energy consumption. Therefore, longer burst i
200there would be high
of the inactivityzz
timers and thus less energy consumption.
zz
4
How to find the optimal burst size?
•  Proxy does not know client’s TCP receive buffer size
•  Could design a protocol where client informs proxy
server
–  Need custom streaming client software à bad idea
•  Insight: TCP flow control messages indicate too large
burst size
–  Proxy receives TCP zero window advertisements from client
–  Provides transparent way of identifying too large burst size
•  How to probe for right size?
–  Linearly increasing burst size can take a long time
–  Use binary search instead
o/3G/LTE
p/ Wi-Fi
p/3G/LTE
/Wi-Fi
/3G/LTE
16%–328–38
13%–458–38
27%–458–38
15%–452–33
30%–452–33
s
s
s
s
s
14%–328–39 s
–
–
–
–
4%–280–3 s
–
–
–
–
50%-2000-31 s
–
54%–2000–39 s
–
55%–452–33 s
How to find the optimal burst size?
ngs at Topt for di↵erent mobile devices using EStreamer, while playing di↵erent audio and video streami
Initially
d LTE.• The
LTE measurement results are only with HTC Velocity phone.
–  Set max burst size according to
initially buffered data
Burst Interval
•  For each burst
12
8
4
40
Burst Interval
10
zwa = 3
1330
8
–  Set new burst size to (max-min)/2
6
–  Check
whether
received
any
ZWAs
1320
zwa = 3
zwa = 2
•  No à set min=current burst size4
•  1310
Yes à set max=current burst size
2 and
revert back to previous burst size
1300
0
• 500 Stop search
when
0
10
20 max-min
30
40 equals
50
Round
minimum increment
1250
40
30
Burst Interval
–  Start
min=0
=3
1340fromzwa
1500
Power Consumption mW
16
Burst Interval
Consumption
nterval
•  Proxy calculates it during the Fast
1350
Start period
Power Consumption
Power Consumption mW
20
452 kbps Dailymotion video
stream to Nexus One over 3G
zwa=4
1000
20
750
Burst Interval
10
Power Consumption
500
0
2
4
6
Round
8
0
10
E.g. 1s2:worth
of content
Internet Ra-–  Figure
Streaming
128 kbps Internet Ra- Figure 3: Streaming 452 kbps Dailymotion
dio to Nokia E-71 over 3G.
video to Nexus One over 3G.
e name this burst sending event
previous work [5]. Power savings for these two experimen
How much energy can be saved?
•  Overall large savings possible
–  audio: 36%-65%, video: 20%-55%
•  Savings depend largely on network type and parameters
–  3G/LTE have longer inactivity timers than Wi-Fi
–  Parameters (timers and DRX) determine the tail energy that can
be saved
•  Stream rate matters as well
–  Bursting lower rate stream yields larger savings
•  Smaller savings with video streaming compared to audio
–  Display draws significant amount of power
–  Video decoding is more work than audio decoding
Outline
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
–  Traffic shaping in multimedia streaming
–  Optimized prefetching of video streaming
–  Context-aware scheduling of transfers
•  Conclusions
On-demand mobile video stream delivery
encoding rate
Fast start
server throEling
non-­‐persistent ON-­‐OFF
persistent ON-­‐OFF
whole video download
DASH
!me
•  Different strategies to deliver video stream content to client
–  Caused by client software+streaming service combinations
–  Leads to very different energy consumption
Energy consumption
of different
Energy Efficiency
of the Streaming Stra
strategies
•  Several sources of energy
inefficiency
–  Underutilization of the capacity
–  Auxiliary TCP control traffic (ONOFF-S)
–  Tail energy in non-continuous
content reception
Average Current Consumption (mA)
350
LTE
3G
Wi−Fi
300
250
200
150
100
50
0
ON−OFF−MON−OFF−S Throttling
Encoding
Rate
Fast
Fast
Caching Caching(20%)
Figure: Avg. current consumed
by with
Wi-Fi,
3G and LTE
Wi-Fi
PSM-A
•  Fast Caching seems best
3Gstreaming
(HSPA) with
FD
Galaxy SIII 4G using different
techniques.
–  But users typically abandon viewing!
LTE without DRX
–  No longer best when aborting after
I Wi-Fi interface is managed using PSM-A, HS
20%
–  Too aggressive prefetching
à
Dormancy,
LTE without DRX.
download content that
will never be
I Playback
current consumption is around 200
watched
I
I
Wireless interfaces can deplete battery at equ
Fast Caching is the least energy consuming t
if the complete video is watched.
Optimizing content delivery strategy
•  Must strike a balance between two sources of energy
waste:
–  Prefetch content in large chunks in order to minimize the tail
energy
–  Limit chunk size in order to reduce the amount of downloaded
content that will never be viewed
•  Need an estimate of when a user will abandon viewing
the video
à use viewing statistics
rs’ Viewing Statistics of a Video
Video viewer retention
Audience retention
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
Fraction of video
0.8
1
•  Video
clip specific
statistics
about
how a
Figure:
Audience
retention
graphsgive
frominsights
YouTube
videos.
I
I
I
new user would view the video
E.g.: poor
content
à likely toofabandon
early that is still
Each – curve
shows
the fraction
the audience
•  YouTube
suchofstatistics
watching
at a collects
given point
time.
Available
only
to “video of
owner”
Some– retain
clear
majority
audience until the end while
others lose most of the audience at the very beginning.
Video clip specific statistics give useful insights about how
a new user will watch the video.
eSchedule algorithm
•  Algorithm that calculates optimal download schedule
•  Input:
–  Viewer retention data for the video clip
•  Ri: “Which fraction of users viewed until time step i”
–  Power consumption characteristics of WNI used (Prx, Etail)
–  Video stream rate (renc) and bulk transfer capacity (rdl)
•  Output: dl schedule S that minimizes expected value of
energy waste
–  Exp amount of energy spent by downloading content that won’t
be viewed
–  Estimate of total tail energy expenditure
–  S is a concatenation of chunk sizes (n-tuple): S=(T1, T2, …, Tn)
X
energy depends on the chunk
size because⇥ the inactivity
ofj =video
Spd
for
Tmin
tolength.
i step
Tmin
thetimer
totalvalues
tailprobability
energy
and that
energy
(i
k)
T ⇥ renc
k ⇥data
renc (2)
schedules,
even
for
a video
can
be+ downloading
directly
computed
abd
wing sessions conditional
corresponding
to
apspent
specific
radio
interface
l
m
Ecurrent = Emin (i j) +
one minute in
video
hastom
presented
[13]
k=
+1 is Eq.
can
T . The
tailWe
energy
added
chunk
which
isaudience
not watched.
obtain
(1) for
for afor
calculating
if Ecurrent < Emin
(i) then
theexceed
retention.
The
probability
a user to(e.g.,
will watch the from
sible
schedules).
if the value
user abandons
viewing
in the downloading
middle of
Emin (i) = Ecurrent
;
computational
comp
theeven
expected
of energy
waste
when
iour depends abandon
during
a
discrete
time
step
j
measured
from
the
However,
we
can
use
an
lastchange(i)
=
j;
In
the
above
equation,
r
is
the
encoding
rate
of
enc
downloading it. When downloading
a chunk containing
a
video
chunk
corresponding
to
T
seconds
of
content
is
the
number
of
sm
dynamic
= D programming whic
engaging the beginning
streaming.
We
needthe
theis
probability
of as
the
user end
T the
seconds
worth
of also
content,
expected
value
of follows:
the
of the
video
viewing
computed
while
end
0 do subproblem
on >
solving
abandoning
viewing
(pabd ).content
This i.
probability
needs
be based
unnecessarily
downloaded
isIncalculated
using
starting
from
discrete
time
step
this case,
Tto
repvideo
can
be divided
the potential
Probability
that
user
interrupts
during
a
discrete
time
step
j:
S
=
(end
for reuse in lastchange(end),
later iterations.
re-evaluated
eachconsists
time the
next
chunk
sizefirst
is chosen
(2).
This
equation
of
two
parts:
the
one
end
lastchange(end);
an scheduler
integer value.
In the equation,
rdl Itisis the
uling to save resents
by=from
the
minimum
chu
is computed
piece
by
during
theatstreaming
session.
a schedule
corresponds
phase during
chunk
viewer
pabdthe
(j) =to the
p(“abandon
j”)which
= Rj the
= 0) computed
1 pj (X
output
S;
at which
contentdownloaded
is that
downloaded
(i.e., computed
TCP
bulk tationalWe
complexity
grows
m
ficant amount rate
conditional
probability
can be
outline
the com
being
simultaneously
anddirectly
consumed,
and
retention
data
(Ri) Specifically
=
R
(1
p
(X
=
1))
=
R
R
(3)
j
1
j
j
1
j
of
video
length.
from
the audience
for abeing
user to
the
second
one to the
phase
which
it is consumption
only
capacity),
Pretention.
(rdl )during
isThe
theprobability
power
rx
not leverage transfer
in Algorithm 1. The
abandon
during
a
discrete
time
step
j
measured
from
the presented in [13] to our pr
consumed.
receiving data at the download
rate,
(T ) chunk
is the
exponentially
the T):
video len
Exp. value of while
unnecessarily
content
for aEtail
given
(starts
at as
i, partitioning
dur
optimal
computational
complexity
of
beginning of downloaded
the video viewing
is computed
as follows:
l
m
limit
the
granularity
of
chunk
amount
of
tail
energy
spent
after
downloading
a
chunk
the
number
of smallest
po
which mainThe download schedule
consists ⇥of a concatenation
ofis second,
the
energy
waste. to
Th
X
five
it divided
is infeasible
⇤
video
can
be
into
(i.e
carrying
seconds
of
content,
and
pabd
(i
⇥j r1dlpE[B
k⇥
+T )]
E[BwasteT
T(j)
)] = = worth
p(i,
p(“abandon
at +
j”)k)=kR
=r0)
waste
enc (i,
ip [11]. These video
j (X
abd
amount
of
buffer
approach
bycontent
comparing
thesize
ene
chunks
that can
be of
different
sizes.
We
denote
aby the
ning
of thein
video
w
minimum
chunk
k=1
=
R
(1
p
(X
=
1))
=
R
R
(3)
j
1
j
j
1
j
is
the
expected
value
of
content
of
the
T
seconds
long
schedules,
even
for
a
video
las
ous feedback download schedule as X
at given
time
instance
T
outline
the
computatio
an
n-tuple S =⇥ (T1 , T2 , T3 , . . . , ⇤T(e.g.,
and
adds
in
each
i
n ) We
one
minute
has mo
chunk starting from time steppabd
i that
is Tdownloaded
and
(i + k)
⇥ renc
k ⇥ renc
(2) in Algorithm 1.video
during video which
The
first
par
l
m having durations T . The exconsists
ofbecause
nk=chunks
schedules).
the partitioning
consideredpoints
pieceo
i
download
schedule
consists
of a concatenation
of sible
optimal
watched
the+1user
abandons
the session.
examples of never The
However,
we
can
use
an a
video
chunks
thatenergy
can be of
different
sizes.
We denote
a the energy
pected
value
of the
waste
of an
entire
download
waste. The
algori
the
optimal
schedul
programming which
os which we Exp.
ofschedule
energy
a
single
Invalue
the above
equation,
is the
encoding
enc for
download
aswaste
anrn-tuple
S=
(T1 , T2chunk:
, T3rate
, . . . ,of
Tn ) dynamic
ning of the video with a sm
schedule
S
can
be
calculated
using
Eq
(4).
mined.
Redundant
on solving
subproblemsc
the
streaming.
We
need E[B
the
probability
userex- based
(i, T )] of Tthe
waste
which
consists
of also
n chunks
having
durations
ys starts from
and
adds
in
each
iteration
i . The
tail
energy
+
⇥
P
(r
)
(1)
E[E
)]
=
E
(T
)
+
for
reuse
in
later
iterations.
In
waste (i, T viewing
rx
tail
dl
abandoning
(p
).
This
probability
needs
to
be
abd
optimal
schedule
pected value of the energy
wasterof
the the
considered
piece
of vid
dl an entire downloadunnecessary
, the fraction
dl
energy
is computed piece by
re-evaluated
eachbetime
the next
chunk
size is chosen schedule
schedule S can
calculated
using
Eq (4).
the optimal
schedule for
th
beginning
E[E
(1,
S)]
=
E
(T
)
+
tational complexityuntil
growstime
mu
waste
1the streaming session. It is a
tail
iven point of
by
the
scheduler
during
Redundant calculatio
Exp. value of energy waste
for entire dl schedule: of mined.
video
length.
Specifically,
w
j 1be directly computed
conditional probability
that P
can
in
the
subsequent
ite
T
l
the
optimal
schedule
(E
(i)
n
E[Ewaste (1, S)] =XEhtail (T1 ) +l=1
min
i
X
presented
in [13]
our prob
discount future
tailtoenergy
from the audience retention. The
P probability for a user to
beginning
until
time
step
i)
the
identified
partit
T p
n
1
(k)
⇥
E
(T
)
+
i j computational
abd
tailthe
X h time step
X j measured
that may notcomplexity
happen of O
abandon during a discrete
from
T ⇥renc
rdl
⇤
eSchedule algorithm: problem formulation
T ⇥renc
rdl
T ⇥renc
rdl
j 1
l=1
l
subsequent
iterations.
1
pabd (k) ⇥ Etail (Tj ) + is in
thethe
number
of smallest
poss
the
schedule
consisti
j=2 viewingk=1
beginning of the video
is computed
as follows:
the
identified
partitioning
j=2
k=1
video can be divided into (i.e., vp
n
i
h
chunks
tochunk
download
j 1
XX
n h⇥
consisting
of Tth
⇤ ⇤i PrxP(r
⇥ waste X
bythe
theschedule
minimum
size
(r)
m
rxdl
dl )
pabd (j) = p(“abandon
at
j”)
=
R
p
(X
=
0)
j
1
j
P
) ⇥
(4) (4)
E EBBwaste
chunks
to download.
j 1 ( (TT
jk), Tj ⇥
We
outline
the
computation
T
r
k
k=1
= Rj j=1
(1 pj (X
= 1))
Rjrdl dl (3)
k=1= Rj 1
j=11
in Algorithm 1. The first part it
3.4
eSchedule-h:
partitioning
points
ofH
The energy optimal schedule is the one that minimizes optimal
3.4 eSchedule-h:
Heuristic
The
download
schedule
consists
of a(4).
concatenation
of the energy waste. The algorith
energy
waste
asschedule
calculated
Thethe
energy
optimal
iswith
the
one
that minimizes
We
also
designed
ah
We of
also
a heuristic
video chunks that can be of different sizes. We denote a ning
thedesigned
video with
a smal
the energy
waste
as calculated
(4).
h that
can
used
download
schedule
as an n-tuplewith
S = (T
and
adds
in be
each
iteration
h that
can
be iterative
useda
1 , T2 , T3 , . . . , Tn )
3.3
Scheduler
Using
Dynamic
Programming
of
the
schedule.
The
heurist
which consists of n chunks having durations Ti . The ex- the considered piece of video
video chunks that can be o
download schedule as an n
which consists of n chunks
pected value of the energy
schedule S can be calculate
eSchedule algorithm: solving optimum
•  How to find schedule S that minimizes E[Ewaste (1, S)] ? =
•  Brute force search does not scale
–  Number of possible solutions grows exponentially with video
length (1 min video à over 2 M possibilities)
•  Use dynamic programming
Etail (T1 )
n h
X
j=2
n h
X
1
⇥
E B
B. Jackson et al., “An algorithm for optimal partitioning
of data on an interval,” Signal Processing Letters,
j=1IEEE,
vol. 12, no. 2, pp. 105–108, Feb 2005.
–  Iterative algorithm
–  Compute and store solutions to subproblems The energy optimal sched
•  Optimal schedule for a part of the video
the energy waste as calcula
–  Use stored solutions in subsequent iterations
•  Avoid redundant calculations
3.3 Scheduler Using Dyn
–  Feasible complexity: O(n2) where n is videoEach
length
in mincan
sizebe of an
chunk
chunks
fined minimum to the rem
sequently, the number of
Performance evaluation
•  Simulations using Matlab
•  Short and long videos
–  5 min threshold
–  16 videos in total
•  Retention and other data from real YouTube videos
•  Abandoning time randomly drawn from the probability
distribution given by retention
•  Power models based on measurements with Samsung
Galaxy S3 LTE
•  Video download schedule computed using eSchedule and
total energy consumed was calculated for each session
–  Compare to ON-OFF and whole video download
Performance evaluation
•  ON-OFF strategy is fairly
good when tail energy is
small
–  Wi-Fi and LTE with DRX
•  Whole video dl pretty good
with short videos and large
tail energy
•  eSchedule roughly halves
the energy overhead
short
videos
long
videos
compared to oracle (always downloads
the right amount in one shot)
Android app
0.5
0
0
0.2
0.4
0.6
0.8
fraction of video watched (%)
(a) Wi-Fi
0
1
140
120
1
0.9
3G
0.8
100
0.7
80
0.6
0.5
60
0.4
40
0.3
0.2
20
0
0
audience retention
Wi-Fi
energy waste compared to oracle (J)
10
StreamThrottler
YouTube App
1
Fast Caching
ON−OFF−M
audience retention
audience retention
energy waste compared to oracle (J)
•  Implemented
a prototype app for Android
Android
App :also
StreamThrottler
0.2
0.4
0.6
0.8
fraction of video watched (%)
1
(b) 3G
Figure: Estimated energy waste compared to Oracle while streaming
a 399s video using Galaxy SIII 4G.
Outline
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
–  Traffic shaping in multimedia streaming
–  Optimized prefetching of video streaming
–  Context-aware scheduling of transfers
•  Conclusions
Context-aware scheduling
•  We’ll look at Bartendr as an example
–  Aaron Schulman et al. Bartendr: a practical approach to energyaware cellular data scheduling. In ACM MobiCom 2010.
–  Learning to schedule transfers in suitable times
How signal strength impacts energy?
Traffic Scheduling
•  Two-way impact: throughput and transmit power
HSPA, phone1
HSPA, phone2
LTE, phone2
3000
2000
1000
12
10
throughput (Mbit/s)
power draw (mW)
4000
8
6
4
2
0
0
6
12 18 24 30 36 42 48 54 60
signal attenuation (dB)
0
0
6
12
18
24
30
36
42
48
(a) Signal strength vs. power for one (b) Box plots of signal strength vs. throughput
LTE/HSPA and one HSPA only smartphone for two di↵erent HSPA phones
Figure 12.9 Impact of signal strength on power draw and throughput
Bartendr: basic idea
•  Moving phone experiences varying signal strength
•  Idea: Schedule transfers to happen at times of good
signal strength
–  Save energy by holding those transfers that are not time critical
•  Example applications
–  Background sync: 5 min interval sync could be more efficient if
done sometime between 4 to 6 min
–  Streaming media: Consume buffer when the signal is weak,
prefetch when the signal is strong
•  Challenge: How to know when to transmit and when to
hold?
–  Need prediction
Bartendr: signal tracks
•  People tend to move along same paths in their day-today life
•  Bartendr predicts signal strength for a phone moving
along a path
–  Use previous signal measurements captured while traveling
along the same path
–  Signal strength measurements are basically energy free
•  phone needs to do it anyway for handoffs
•  Assumptions:
–  Phones can store several such signal tracks (frequently traveled
paths)
–  Phone can infer currently traveled track using mobility prediction
Bartendr: signal strength prediction
•  Step 1: find the current position of the phone on the
current track
–  GPS draws too much power
–  Find position in track with closest measurement to current one
–  May have many similar strength positions à use also neighbor
base station list
•  Step 2: predict signal in the future starting from current
position
–  Look ahead in previous measurements
–  Speed may differ à continuous update of current position on
track
Bartendr: scheduling bg sync
•  Schedule next sync based on predicted signal strength
•  Sleep in between schedule and sync
eventswhen
à may
make
Scheduling
to sync
Wake-up, sync, schedule, sleep
errors
Uses threshold for efficient sync
•  Two threshold-based schemes:
Schedules for either first or widest signal
widest
–  First above threshold
–  Widest above threshold
•  Comes later but has
larger margin for error
signal strength (RSSI)
•  Comes soon à error small
first
-60
-65
-70
-75
-80
-85
-90
-95
-100
0
50
100
150
200
time (s)
•  Simulations suggest 10% energy savings for email sync
–  Widest outperforms first
250
13
Bartendr: scheduling stream transfers
•  Differs from syncing because continuously awake à can
compensate for speed variations in real time
•  Results into rather similar optimization problem than we
looked at with eSchedule
–  Chunked downloading
–  Energy spending is function of tail energy and energy spent for dl
•  dl energy varies with signal strength
–  Can be solved using dynamic programming too
•  Simulations suggest 60% energy savings
LoadSense
•  LoadSense is kind of follow-up work to Bartendr
–  Abhijnan Chakraborty et al. Coordinating cellular background
transfers using loadsense. In ACM MobiCom 2013.
•  Solution for scheduling background transfers
•  Takes explicitly into account load in the cell
–  Link quality alone is insufficient
–  Load can be passively inferred through power ratio
•  Sense total power in the cellular channel and compare it with the
power of the pilot signal transmitted by base station
•  Schedule bg transfers when sensed load is small
–  Improve energy efficiency by transmitting with higher throughput
Outline
•
•
•
•
•
Intro
Standard power management techniques
Beyond standard techniques
Example optimizations
Conclusions
What else can be done?
•  Smarter (cooperative) scheduling to reduce contention
–  Random access channels (Wi-Fi) can cause energy waste by
idle waiting
–  Multiple clients of same AP or multiple APs in range (same ch)
•  Leverage alternative low-power radios
–  E.g. Zigbee or Bluetooth in conjunction with Wi-Fi
–  Idea is to always use lowest power radio for the job
–  Discovery of Wi-Fi access points using Bluetooth contact
patterns
•  Saves energy spent by Wi-Fi scanning
–  Data transmission with highest possible energy utility
•  Synchronize background transfers by different apps
–  Amortize tail energy by batching bg transfers
61
Summary
•  Energy efficiency of wireless networking
–  Depends on wireless access network technology used
–  Depends on context à signal quality and network load
•  Standard power management techniques
–  Wi-Fi has PSM
–  Cellular network technologies have RRC
–  Tail energy can be mitigated in cellular nws by using
•  Fast Dormancy (3G)
•  Discontinuous reception (LTE and 3G)
•  Optimizing energy efficiency further
–  Traffic scheduling
•  Traffic shaping
•  Context-aware scheduling
–  Smart use of wireless network interfaces
•  Use always the most energy efficient one
–  Application specific optimizations
•  Mitigate mismatch between power mgmt and application behavior