Download Slide - Interactive Computing Lab

Energy Management: Part I
Uichin Lee
KAIST KSE
Mobile Processing Power –
Changing the Mobile Device
From http://www.ieee-infocom.org/2010/docs/Infocom2010_keynote.pdf
Chipset Business Evolving to System
Business
Integration is key to driving advanced functionality to mass market
From http://www.ieee-infocom.org/2010/docs/Infocom2010_keynote.pdf
Creating New Mobile, Computing and
CE Device Categories
From http://www.ieee-infocom.org/2010/docs/Infocom2010_keynote.pdf
But Major Gaps Exist
Battery Technology is Falling Behind
How do we balance battery life with performance and cost?
Energy Management
on Handheld Devices
Marc A. Viredaz, Lawrence S. Brakmo, William R. Hamburgen
HP Labs ACM Queue Oct. 2003
Itsy Platform
• Itsy goals (YR 2000):
– Small, powerful, flexible h/w platform
– Flexible, extensible, advanced s/w environment
• Base system
– StrongArm SA-1100 microprocessor
– 32MB DRAM & Flash Drive
– LCD display and touch screen
• Passive matrix gray scale
–
–
–
–
–
Li-ion battery (charges from USB)
2-axis accelerometer
Microphone
Jacks (headset, docking)
Daughtercard connecter
• Software
– Linux 2.0.30 w/ modified memory/flash-based file systems
– Power management capability
Ohm’s law, Power, Energy
• Ohm’s law: V=IR (=current*resistance)
I=V/R?
9v
2.2Ω
Voltage
drop = 16v
2kΩ
I=V/R?
• Power: watt (W) = 1 joule/second (J/s)
– Power (W) = VI= I2R
• Energy: (Ws, or Joule)
– Energy (Ws) = power (w) * time (s)
Power Measurement
• Voltages are directly measured.
• Currents are calculated from the
corresponding sense-resistor voltage drop.
• Elementary power domains are delimited
by dashed lines.
Power Consumption
Energy Saving Techniques
• Simple approach: if a unit is not used, turn off
or put into sleep mode
• But requires well structured h/w and s/w
design
• Inter-connected building blocks must independently
function and be independently powered on/off
• Operating systems (or applications) utilize measured
power values to balance performance and battery life
• Major power draws: processor, memory,
display, audio system, wireless networking
Processor
StrongARM SA-1100
Processor
• Frequency or/and voltage scaling: Power ∝ f·v2
– f: operating frequency, v: voltage
Voltage Scaling in Strong Arm SA-1100
59Mhz at 0.79v
 251Mhz at 1.65v
J.Pouwelse, K.Langendoen, and H. Sips, “Dynamic Voltage Scaling on a Low-Power Microprocessor”, MOBICOM2001
Processor: How to Scale?
• How to scale voltage/frequency?
power
Low frequency
?
High frequency
...
time
Watts
Watts
Eactive
Eactive
Eidle
t
t
Critical Power Slope: Understanding the Runtime Effects of Frequency Scaling, ICS2002
Processor: Critical Power Slope
m  mcritical
mcritical  P
f min
Power
f
: energy efficient to run at lower freq
 Pidle
min
mcritical
Pf min
m  mcritical
: energy efficient to run at higher freq
Pidle
f
min Frequency
fmin: min operating frequency
Pfmin: power consumption at freq fmin
Pidle: idle power consumption
Critical Power Slope: Understanding the Runtime Effects of Frequency Scaling, ICS2002
Memory
• DRAM typically has four states:
– Activate/pre-charge: read/write happen (most energy consuming)
– Fast lower-power: short-term sleeping (w/ fast wake-up time: ~10ns, and
consumes only half of the active power)
– Self-refresh: only refreshing is happening (much less power consumption,
requires several 100 cycles)
– Deep power-down: refreshing stopped (lost data)
Figure from: http://research.microsoft.com/pubs/102932/flicker-tr-2009.pdf
Display
• LCD itself consumes minimal energy, yet display
front- and back-light dominates..
– Possible to dim lights of “light” pixels (for energy saving)
• Organic light-emitting diode (OLED)
– Better quality than LCD (fast response), but it’s emissive
and can’t make use of ambient light (energy consuming..)
Wireless Networking
• Technologies
– WiFi, Bluetooth
– 2G/3G/4G cellular communications
• Power consumption: BT < WiFi < 2/3/4G
• Caveats:
– bit/joule must be considered
– bit/joule varies with data rate
Cool-Tether: Energy Efficient On-the-fly WiFi Hot-spots using Mobile Phones, CoNext 2009
Decomposing power measurements
for mobile devices
Andrew Rice and Simon Hay
Percom 2010
We want to know how much energy
a particular action will consume
Example: joining the wireless
network consumes 6 Joules
HTC G1 (or Magic), Android 1.1, 194 trials
We measure energy consumption
by intercepting the power supply
Both voltages
are sampled
at 250 kHz
V1
0.02Ω
V2
Power ∝ V1 x V2
Trace of the G1 boot process
HTC G1 (or Magic), Android 1.1
Joining a wireless network: DHCP
• Dynamic Host Configuration Protocol (DHCP)
– Provides automatic configuration of the host connected to network
– Provides hosts with initial configuration information upon bootup:
• IP address with subnet mask, default gateway, IP address of the DNS server
server A
server B
client
(not selected)
DHCP discover
(selected)
DHCP discover
determine
configuration
determine
configuration
DHCP offer
DHCP offer
select configuration
DHCP request
DHCP request
DHCP ack
Initialization completes
using the allocated configuration
graceful shutdown
DHCP release
discard lease
Joining a wireless network: ARP
• Address Resolution Protocol (ARP) and
Reverse ARP (RARP): translation between IP
and MAC addresses
TCP
IP address
(32 bit)
ARP
RARP
Ethernet MAC
address
(48 bit)
UDP
ICMP
IP
IGMP
ARP
Network
Access
RARP
Media
Transport
Layer
Network
Layer
Link Layer
Joining a wireless network: ARP
Argon
128.143.137.144
00:a0:24:71:e4:44
Router137
128.143.137.1
00:e0:f9:23:a8:20
ARP Request:
What is the MAC address
of 128.143.71.1?
Argon
128.143.137.144
00:a0:24:71:e4:44
Router137
128.143.137.1
00:e0:f9:23:a8:20
ARP Reply:
The MAC address of 128.143.71.1
is 00:e0:f9:23:a8:20
Access point beacons correlate with
spikes in the power trace
HTC G1 (or Magic), Android 1.1
Timestamped events from the
phone must be aligned with the
appropriate sample points
The synchronization information is
embedded in power trace
Bright screen
Dimmed screen
HTC G1 (or Magic), Android 1.1
Hypothesis matching pulses
g(t)
HTC G1 (or Magic), Android 1.1
Find alignment from autocorrelation
with a hypothesized signal
autocorrelation (f*f)(t) – cross-correlation (f*g)(t)
f(t)
Cross-correlation: a measure of similarity of two waveforms as a function
of a time-lag applied to one of them.
Autocorrelation: cross-correlation of a signal with itself; there will always
be a peak at a lag of zero, unless the signal is a trivial zero signal.
HTC G1 (or Magic), Android 1.1
HTC G1 (or Magic), Android 1.1
Remove the DHCP overhead by
using static addressing
HTC G1 (or Magic), Android 1.1
Static addressing reduces the
connection cost to 1.5 Joules
Static Addressing
Dynamic Addressing
HTC G1 (or Magic), Android 1.1, Static = 143 trials, Dynamic = 194 trials
We could remove the ARP probes
from our client implementation
• “ARP probe”: ARP probe is broadcast to see if
the address is already in use
• RFC2131 “...the client SHOULD probe the
newly received address, e.g., with ARP.”
• RFC2119 – SHOULD “...there may exist valid
reasons in particular circumstances to ignore a
particular item”
Enter the
Android 2.1 doesn't ARP probe in
our tests
Dynamic addressing now costs 1.5J
Dynamic Addressing N1
Dynamic Addressing G1
Google N1, Android 2.1, 100 trials / HTC G1 (or Magic), Android 1.1, 194 trials
Joules consumed
The G1 histogram peaks are due to
discontinuities in connection time
Time to connect (seconds)
HTC G1 (or Magic), Android 1.1, Dynamic
HTC G1 (or Magic), Android 1.1, Dynamic
DHCP Finish
DHCP Finish
DHCP Start
DHCP Start
Power (Watts)
Power (Watts)
Caused by power control in radio?
This power control is evident when
sending data too
Send 7K of data over TCP
HTC G1 (or Magic), Android 1.1
Send 8K of data over TCP
This effect has a big impact on
energy cost
HTC G1 (or Magic), Android 1.1, 1120 Trials (HTC Hero, Android 1.5 is the same)
N1 energy performance
Best case: same
Google N1, Android 2.1, 900 Trials
Worst case: much better