Wireless Network Interface Energy Consumption of Popular Streaming Formats

1
Wireless Network Interface Energy Consumption of
Popular Streaming Formats

• Goal Conserve client WNIC energy consumption for
  popular streaming formats
• MS Media, Real and Apple Quicktime
• Use traffic shaping network proxies to
  aggressively transition the WNIC to lower energy
  consuming sleep state
• Show the possibility for significant energy
  savings

Surendar Chandra University of Georgia http//gree
nhouse.cs.uga.edu/
2
Motivating Scenario

• Commodity mobile devices and ubiquitous wireless
  networks
• Cooperating users in a sports stadium
• stadium provides network infra-structure
• users provide resources for ad-hoc peers
• prototype being deployed at UGA
• Passive participants
• access objects provided by the infra-structure
• e.g scores, game replays
• Active participants
• Create content for fellow peers
• e.g. fan commentary, views from any angle

3
Issue 1 Content adaptation
Original JPEG 116 KB

• adapt content for the particular mobile devices
  resource constraints
• My Ph.D. thesis image transcoding
• Developed quality aware image transcoding
• predict expected size savings, computational
  overhead for transcoding and image quality factor
  loss
• Applications
• Slow networks - web proxies
• QoS - busy web servers
• Battery storage - digital camera

Low JPEG Quality 10 KB
4
Issue 2 Battery conservation (todays focus)

• Continual improvement in device capabilities
• faster, more memory, better color displays
• Battery capacity improvement slower
• Continual pressure for further device
  miniaturization

Processor Performance (Moores Law)
Battery Capacity (Evereadys Law?)
Courtesy of Ravi Subramanian (MorphICs)
5
Typical energy consumption values for iPAQ

• Note
• energy consumed depends on the components and
  their energy states
• iPAQ battery capacity 2950mAh (22850mWh _at_ 3V)

Source Compaq researchers, Sukjae Cho, Paul
Havinga, Mark Stemm
6
Research Goal

• Goal consume energy proportional to stream
  quality
• users choose the stream quality based on energy
  budget
• energy consumption for consuming and producing
  content
• Topic for today consuming multimedia
• popular streaming formats (MS media, Real,
  Quicktime)
• develop policies to reduce energy consumption
• Technique Aggressively transition WNIC to a
  lower energy consuming sleep state (instead of
  active idle)
• Only analyze energy implications for the
  multimedia player

7
Outline for the rest of the talk

• Motivating the general research focus and the
  specific research problem
• reducing energy to view popular multimedia
  streams
• Explore IEEE 802.11b power saving mode
• Cooperative proxy infra-structure
• Experimental setup
• Key results
• Conclusions
• We can save significant energy on the clients
• Future directions
• Energy efficient mechanisms for serving
  multimedia
• Energy aware peer-to-peer overlay networks

8
IEEE 802.11b power saving mode

• Scheduled rendezvous mechanism using beacons
• AP informs wireless station of pending packets
  at predefined intervals (beacon interval)
• Clients transition to lower energy consuming
  states between scheduled beacons
• Energy savings depends on Wait intervals

beacon
1
1
2
Access point
beacon with data
Beacon Interval
Wait Interval
PS poll
1
sleep
PS data
Wireless station 1
9
Significance of Wait interval

• Simulation results to highlight Wait interval
• Small changes can significantly affect energy
  consumption

• Wait intervals over 100 msec would trigger
  clients to adapt to lower quality

10
Experimental results

• Carefully opened a WNIC card, hooked up probes to
  measure the efficacy of PSM mode
• WNIC continously stays in high energy state for
  streams over 56 Kbps

High energy state
WNIC energy state
WNIC data reception
11
Energy savings depends on Wait interval

• IEEE 802.11 does not define bounds for this
  interval
• Small interval
• Good for isochronous streams (semi-regular with
  real-time constraints)
• But, tends to give higher priority to clients in
  PSM
• Multicast packets do not need PS poll. AP
  directly sends the data packet right after a
  beacon
• Access points typically discourage multicast
  packets
• IEEE 802.11 PSM designed for asynchronous and
  infrequent data e.g. login session
• Does not understand requirements of isochronous
  streams

12
Towards stream aware traffic shaping

• Goal Develop mechanisms to aggressively
  transition the WNIC to lower energy consuming
  sleep state
• Step 1 Understand stream dynamics for popular
  media formats
• Microsoft media, Real, Apple Quicktime
• More likely to be deployed than research formats
  specifically designed to conserve energy
• Step 2 Explore energy conserving techniques
• Client-only history based policy to predict
  inactivity periods
• Assistance from traffic shaping network proxies

13
Step 1 Experiment setup
Browsing station
Multimedia Server
Traffic Shaper (dummynet)
Access Point
Monitoring Station (tcpdump)

• Digitized movie trailer
• 11 Mbps 802.11b WLAN
• Multimedia service
• Microsoft media service (Win 2000 Server)
• Realserver 8.0
• Apple Darwin Server
• Dummynet traffic shaper simulates lossy network

14
Stream energy characteristics

• Negligible difference in energy consumption even
  though orders of magnitude difference in data
  consumed for different streams
• Idle and read power states of WNIC consume
  similar energy

15
Microsoft media stream characteristics

• Packets arrive at fairly regular intervals
• Can assist client side energy conserving policies
• Large packets (up to 16 KB) Uses network
  fragmentation
• Losing one fragment loses the entire packet
• Fragments arrive back to back
• can assist adaptation policy
• Lossy network - adapts to a low quality stream
• fragmentation makes this effect worse

16
Real stream characteristics

• Variable packet size (50-1500 bytes)
• Packet arrivals almost regular
• Packets sent closer to each other
• Packet size less than MTU
• No network level fragmentation
• Lossy networks - lower quality video

17
Quicktime characteristics

• Variable packet size
• Uses multiple ports audio, video streams
  separate
• Packets sent in clusters burst and extended
  idle
• Application level fragmentation?
• Simple adaptation policies need to understand
  this protocol behavior
• Packet size less than MTU
• No network fragmentation
• Lossy networks lower quality video

18
Energy saving potential Idle slots

• Streams spend significant time waiting for data
• Potential for considerable energy saving if we
  can predict future idle slot

19
Energy saving potential Idle slots

• Streams spend significant time waiting for data
• Potential for considerable energy saving if we
  can predict future idle slot

20
Step 2a. Client Side Adaptation Policies

• Use history to predict the arrival time of next
  packet and transition WNIC to sleep state
• If conservative lost opportunity
• If too aggressive and packet has already started
  miss the whole packet
• Simple policies necessary for these mobile devices

Conservative prediction
Aggressive prediction
Decision point
Packets
time
21
Client-only Adaptation Policy Parameters

• History depth
• Number of past events
• Prediction threshold
• conservative level
• Prediction avg. history - threshold
• Mis-prediction back-off
• If we missed a packet, we dont know if a packet
  arrived sooner or never arrived

22
Energy Savings
Orig. Recv Energy (J)
Adaptive energy (J)
Stream Format
Adaptive data loss ()
B/W (Kbps)
158
35
1
56
Microsoft Media
160
60
0.5
256
175
135
0.15
2000
119
55
10
56
Real
124
120
5
256
150
41
30
56
Quicktime
149
47
23
256

• Microsoft Media can offer significant savings

23
Energy Savings
Orig. Recv Energy (J)
Adaptive energy (J)
Stream Format
Adaptive data loss ()
B/W (Kbps)
158
35
1
56
Microsoft Media
160
60
0.5
256
175
135
0.15
2000
119
55
10
56
Real
124
120
5
256
150
41
30
56
Quicktime
149
47
23
256

• Apple and Real are harder to predict lose
  significant data

24
Summary checkpoint

• Step 1 Explored the WNIC energy implications of
  receiving multimedia streams
• MS media packets regular with network
  fragmentation
• Real packets regular and closer together
• Quicktime packets clustered
• Step 2a Simple client-only policies offer energy
  savings for MS media
• 28.8 Kbps MS media 80 energy saving (2 data
  loss)
• 768 Kbps MS media 57 energy saving (0.3 data
  loss)
• Step 2a Real and Quicktime are harder to predict
  
• Increased data loss for energy savings

25
Step 2b. Traffic shaping network proxies

• Network Infrastructure conditions the packets to
  arrive at predefined intervals
• Can also reduce contention among multiple clients
  consuming different streams

26
Architecture
Local Proxy
Client Side Proxy
Multimedia Server
Mobile client
Access Point

• Same parameters as Step 2a MS, Quicktime and
  Real servers on Windows 2000 server
• Local proxy informs CSP on the beacon interval
  for the specific client (fixed intervals
  explored)
• Measure performance in energy metric
  (Joules/Byte)
• Sometimes, clients adapted to the introduced
  delays to lower quality. Energy metric a
  normalized mechanism to measure performance

27
Energy metric for 1 client - Microsoft media
Increasing the delay provides better savings for
lower bandwidth clients Increasing delay for
high bandwidth streams forces an adaptation
For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
28
Energy metric for 1 client - Real
Increasing the delay provides better savings for
lower bandwidth clients Real adapts to delays
quicker
For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
29
Energy metric for 1 client - Quicktime
Increasing the delay provides better savings for
lower bandwidth clients Quicktime adapts for
100 msec delays
For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
30
Energy metric for 2 clients Microsoft Media
Still get savings for two clients
For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
31
Energy metric for 2 clients - Real

• Savings for multiple clients
• Real has trouble with these delays
• we are exploring smoothing client side proxies

For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
32
Energy metric for 2 clients - Quicktime
Still get savings for two clients
For a given stream, lower energy metric
(Joule/Byte) is better Higher bandwidth streams
have lower energy metric
33
Summary

• We showed the restrictions for IEEE 802.11
  scheduled rendezvous mechanisms for isochronous
  streams
• Step 1 We analyzed the stream characteristics
• MS media regular, Real sent close together
• Step 2a We showed the potential and restrictions
  for client only adaptation policies
• Quicktime and Real lead to heavy data loss
• Step 2b We showed that a traffic shaping proxies
  can offer energy savings without any data loss
• Clients adapt to this introduced delays (mistaken
  for network congestion)

34
Future work Energy efficient serving

• Multimedia devices can service multimedia streams
• Video cameras available for iPAQ
• Weve ported Apple Quicktime server to an iPAQ
  running Linux and can serve the reference streams
  described
• Multimedia servers spend energy transmitting data
  streams and receiving stream feedback from
  clients
• Developing scheduled rendezvous mechanisms
• System parameters
• Clients sharing the same access point
  contention
• Dedicated access points for server and client
• wired connection between APs
• Ad-hoc networking

35
Issue 3 Peer-to-peer resource sharing

• Application level p2p overlay
• provide path connectivity
• can reach any peer from any other peer
• overlay depends on configurable parameters
  (available node energy, network bandwidth etc.)
• Goal Minimal path connectivity in a dynamic
  network