August 26, 2004Annie L. Luo, Carnegie Mellon University
Proxy Supported Power Optimization for Heterogeneous Wireless Clients
Annie L. Luo
IBM T.J. Watson Research CenterMentor: Marcel Rosu
Manager: Chandra Narayanaswami
August 26, 2004Annie L. Luo [email protected]
Power Optimization for Wireless NIC
Active power consumption in WLAN interface: 5-10% in notebooks, 50-90% in PDAs
Existing power-reduction approaches for WLAN clients: 802.11 Power Saving Mode – limited power saving during active
transmissions MAC level – extending sleep time Transport level – energy efficient protocols The unpredictability of incoming traffic causes waste of power
Our approach – Power Aware Web Proxy (PAWP), using: A web proxy to shape HTTP traffic going into client’s WNIC
Based on: Application domain knowledge MAC level configuration Network conditions
August 26, 2004Annie L. Luo [email protected]
Analysis of Incoming WLAN Traffic
TbeaconTlisten
Pow
er
Time [beacons]
Pdoze
Pawake
Pdoze
Pawake
Pow
erN
DA
TA
Tbeacon
123456789
Tlisten
Time [beacons]
ND
AT
A
123456
Packet arrival at AP
Packet arrival at Proxy b) Proxy (Ttimeout=25ms)
Packet arrival at AP
a) Direct (no proxy) (Ttimeout=100ms)
Data transferTimeout
Sources of wasted energy in WLAN interfaces
Proposing: Power Aware Web Proxy that modulates WLAN data
transmission into intervals of high and no traffic
August 26, 2004Annie L. Luo [email protected]
Comparison on Power Consumption
DIRECT
PROXY
ACPIspec.pdf eBay.com
www.eBay.com
August 26, 2004Annie L. Luo [email protected]
Architecture of PAWP (summer 2003)
Client Side
Threads
Fetching
Threads
WL
AN
C
lients
Web
S
ervers
Blackboard
OracleRules
Information perClient
Information perRequest
August 26, 2004Annie L. Luo [email protected]
Limitations in PAWP-2003 Implementation
Simplistic Blackboard: Static data structure Inefficient, linear searching
method and data access No memory management
mechanism
The resulting proxy: Handles only one client at a time Has to be restarted after
downloading each page
August 26, 2004Annie L. Luo [email protected]
A More Realistic WLAN Scenario
Multiple, heterogeneous wireless clients, varying in: Power characteristics Connectivity and
configuration Processing capability Display size, etc.
Clients have different requirements in power management
The proxy needs to treat clients differently
August 26, 2004Annie L. Luo [email protected]
PAWP-2004 Implementation
Improvements since summer 2003 Handling cookies Pipelining requests and responses Enriched decision rules set for efficient traffic shaping
Improvements in summer 2004 Re-implemented Blackboard to maintain information using dynamic
data structures – enables handling multiple clients Used hash table for efficient searching and management of web
objs – increasing proxy speed Added memory management – more realistic proxy behaviors Amount of work: 3K lines of code
August 26, 2004Annie L. Luo [email protected]
Experimental Testbed
Oscilloscope(VellemanPC S64i)
Data collection PC
Digital Multimeter(HP3458A) R=0.53
Vdd=3.3V
+ -VR
Wireless Client(IBM ThinkPad)
Intersil PRISM3 PC Cardwith Extender
Power measurement environment for wireless client network interface card
HTTP protocol trace collection using IBM PageDetailer
- Downloading time distribution- Information of web objects- HTTP headers
August 26, 2004Annie L. Luo [email protected]
Improving the Experimental Testbed
Limitations in the prior-summer’04 experimental testbed Manual operations needed for everything
• Start PageDetailer –– start proxy –– start oscilloscope data collection –– start browser –– clean browser cache –– download web page –– close browser –– stop oscilloscope data collection –– stop proxy –– record PageDetailer information –– Start next round……
Tedious, time-consuming, error-prone More importantly: long intervals between experiments make
measurements inaccurate due to real-time network dynamics
August 26, 2004Annie L. Luo [email protected]
Improving the Experimental Testbed
Limitations in the prior-summer’04 experimental testbed Manual operations needed for everything
• Start PageDetailer –– start proxy –– start oscilloscope data collection –– start browser –– clean browser cache –– download web page –– close browser –– stop oscilloscope data collection –– stop proxy –– record PageDetailer information –– Start next round……
Tedious, time-consuming, error-prone More importantly: long intervals between experiments make
measurements inaccurate due to real-time network dynamics
Summer 2004 improvements: Automation of entire experiment process Coordination between the client, proxy, and data collection PC Implemented using Perl scripts (approx. 1K lines of code) Experiment monitoring made easy, measurement results more
accurate
August 26, 2004Annie L. Luo [email protected]
Complete, Across-The-Board Experiments
(sec)0 30 60 90 120 150 180 210
Direct Proxy1 Proxy2 Proxy3 Proxy4 Direct Proxy1 … …
Based on the new experimental testbed Experiments on each proxy configuration can be done in < 30 sec Quick, automatic switching between configurations Measurements in each set are close in time – avoided deviation
1st set 2nd set
August 26, 2004Annie L. Luo [email protected]
Experimental Results (1)
Website Size [kB] / Num of Objects
ConnectionType
Download Energy [J]
Download Time [s]
Throughput[kB/s]
Inte
rne
t Ex
plo
rer
cnn 285kB/80
DirectProxy
4.781.93 (-60%)
10.46.65 (-36%)
27.4
nytimes 270kB/80
DirectProxy
4.912.17 (-56%)
11.67.57 (-35%)
23.3
washingtonpost 535kB/73
DirectProxy
6.502.98 (-54%)
12.611.0 (-13%)
42.5
msn180kB/85
DirectProxy
6.492.86 (-56%)
15.612.2 (-22%)
11.5
Mo
zilla
cnn 298kB/82
DirectProxy
4.881.52 (-69%)
6.374.72 (-26%)
46.8
nytimes183kB/62
DirectProxy
4.931.23 (-75%)
10.823.71 (-66%)
16.9
washingtonpost504kB/67
DirectProxy
6.722.56 (-62%)
9.748.15 (-16%)
51.7
msn172kB/40
DirectProxy
3.341.30 (-61%)
4.324.11 (-5%)
39.8
August 26, 2004Annie L. Luo [email protected]
Experimental Results (2)
www.nytimes.com Tload [s] DownloadEnergy [J]
no prefetching 5.73 1.82
no response pipelining 6.69 1.92
no request pipelining 5.73 1.71
all features on 5.47 1.63
Cost and Benefits of Proxy Features
Relative energy consumption with Proxy vs. Direct case throughput
Results from WMCSA’04 paper
August 26, 2004Annie L. Luo [email protected]
Contributions
Proposed and implemented a Power Aware Web Proxy Uses application level traffic information to schedule network traffic Enables efficient, client-specific WNIC power management No change required on client side
Explored benefits of improved system predictability Identified possible areas for improvement in browser and web page
design
Publications: MobiSys 2004 poster WMCSA 2004 paper BroadWiM 2004 paper Planning to submit more papers based on work of summer’04
August 26, 2004Annie L. Luo [email protected]
Future Work and Discussions
Future work: Optimize proxy traffic shaping based on various client configuration Proxies for media streaming, VOIP, IM, etc. Proxies for ultrawide band networks
Other research directions utilizing PAWP architecture: Web page rendering Usability analysis and optimization Suggestions on HTTP extensions
August 26, 2004Annie L. Luo [email protected]
Web Page Rendering