1
Weather ForecastingPredicting Performance for
Streaming Video over Wireless LANs
Mingzhe Li, Feng Li, Mark Claypool, Bob
Kinicki WPI Computer Science Department Worcester,
Massachusetts 01609
Presenter - Bob Kinicki

• NOSSDAV 2005
• Skamania, Washington
• June 13-14, 2005

2
Outline

• Motivation
• Experiments
• Tools and Setup
• Experimental Design
• Weather Forecasting
• Weather Prediction
• Weather Predictor
• Weather Maps
• Effectiveness (E)
• Results and Analysis
• Conclusions and Future Work

3
Motivation

• Increasing deployment of streaming multimedia
  over wireless networks.
• The promise of higher wireless link capacities
  (e.g. 54 Mbps with 802.11g)
• Streaming applications may encounter bad wireless
  LAN (WLAN) reception quality due to
• Attenuation, fading, frame collisions, rate
  adaptation
• Contention, MAC layer retries
• A Streaming Users Question
• Can I get good performance here?
• The Streaming Applications Decision
• When should I do media scaling?

• The answer Provide Performance Predictions

4
Outline

• Motivation
• Experiments
• Tools and Setup
• Experimental Design
• Weather Forecasting
• Weather Prediction
• Weather Predictor
• Weather Maps
• Effectiveness (E)
• Results and Analysis
• Conclusions and Future Work

5
Tools and Setup

• Single-level streaming encoded at 2.5 Mbps
• Multi-level streaming with 11 encoding
• levels with maximum level 2.5Mbps
• TCP
• UDP

IEEE 802.11g at 54Mbps
Measurement Tools
6
Experimental Design

• Gauging measurement tool interference
• Baseline experiment CPU usage lt 3
• During measurement CPU usage about 35
• Measurement locations
• Fuller Labs Sub Basement, 1st Floor, 3rd Floor
• Wireless Link conditions
• Good, Fair, Bad
• Number of experiments
• 2 video clips 2 protocols 2 encoded methods
  3 locations 3 conditions 5 times - 10 Bad
  runs thrown out 350 stream runs
• Experimental period
• Winter Break Dec 23-25, Dec 28-29, 2004.

7
Outline

• Motivation
• Experiments
• Tools and Setup
• Experiment Design
• Weather Forecasting
• Weather Prediction
• Weather Predictors
• Weather Maps
• Effectiveness (E)
• Results and Analysis
• Conclusions and Future Work

8
Weather Forecasting
Sky conditions http//www.astro.washington.edu/WW
Wgifs/weather.gif
Probability of snow National Weather
Service   http//www.nws.noaa.gov/
9
Weather Prediction

• Potential Weather predictions
• Average frame rate
• Coefficient of Variation (CoV) of frame rate
• Others
• re-buffer count, buffering time, etc.
• Video Frame Rate Quality Categories
• Good (Sunny) gt 24 fps
• Edge (Cloudy) 15-24 fps
• Bad (Rainy) lt 15 fps

10
Weather Prediction
Figure 4 Cumulative Distribution Function (CDF)
of Average Frame Rate
11
Weather Predictors

• Weather predictors
• Wireless Layer
• Received Signal Strength Indicator (RSSI) (dBm)
• Average Wireless Link Capacity (Mbps)
• Wireless MAC Layer Retry Fraction ()
• Network Layer
• Round Trip Time (RTT) (ms)
• Packet Loss Rate ()
• Application Layer
• Throughput (Mbps)

12
Predictor Analysis

• Figure 2 Average Wireless Capacity versus RSSI

13
Predictor Analysis

• Figure 3 Upstream MAC Layer Retry Fraction
  versus RSSI

14
Weather Maps

• Creating a Weather Map
• Divide prediction
• Good (Sunny), Edge (Cloudy) and Bad (Rainy).
• Put the predictor samples in increasing order.
• Compute prediction probabilities.
• Divide the predictor data into 10
  equally-populated bins.
• Determine the fraction of Good, Edge and Bad per
  bin.
• Draw the weather map.

15
Effectiveness (E)

• Effectiveness (E)
• The fraction of the range of the weather
  predictor in a weather map that is likely to
  produce an accurate prediction.
• The Effective Range, Reffective,, is the range of
  a predictor that provides better than a 50
  chance of yielding a good or bad prediction.
• The Practical Range, Rall , is the useable
  predictor range running from the median of the
  first sample bin to the median of last sample
  bin.
• Thus, 5 outliers are removed from both ends of
  the range to yield the practical range.
• E is between 0 and 1.

16
Outline

• Motivation
• Experiments
• Tools and Setup
• Experiment Design
• Weather Forecasting
• Weather Predictor
• Weather Prediction
• Weather Maps
• Effectiveness (E)
• Results and Analysis
• Conclusions and Future Work

17
Weather Map Analysis Figure 5 Frame Rate
Prediction by RSSI
18
Weather Map AnalysisFigure 6 Wireless Link
Capacity
E 0.97
19
Coefficient of Variation of Wireless Link Capacity
Figure 7 Versus Average Frame Rate
Figure 8 Versus Average Link Capacity
20
More Weather Maps
Figure 9 Upstream Wireless Retry Ratio E 0.75
Figure 10 IP Packet Loss Rate E 0.71
21
RTT Weather Maps
TCP Streaming Videos E 0.83
UDP Streaming Videos E 0.94
22
Throughput Weather Maps
Single Level Encoded Videos E 0.82
Multiple Level Encoded Videos E 0.31
23
Throughput Analysis
Multiple Level TCP Streaming
Multiple Level UDP Streaming
Single Level TCP Streaming
Single Level UDP Streaming
24
Effectiveness Summary
Table 3 Effectiveness of Weather Maps
25
Outline

• Motivation
• Experiments
• Tools and Setup
• Experiment Design
• Weather Forecasting
• Weather Predictor
• Weather Prediction
• Weather Maps
• Effectiveness (E)
• Results and Analysis
• Conclusions and Future Work

26
Conclusions and Future Work

• Reliable performance forecast Predictors
• Wireless RSSI
• Average wireless link capacity
• Regional predictors
• IP loss rate lt 2
• RTT lt 10 ms
• Effectiveness
• varies for different video configurations.
• Single level video performance is easy to
  predict.
• Reliably forecasts of streaming weather can
  benefit video rate adaptation techniques.
• Future Work
• Incorporate prediction into a dynamic video
  system.
• Evaluate prediction with combined weather
  predictors.
• Consider weather maps with different predictions.

27
Weather ForecastingPredicting Performance for
Streaming Video over Wireless LANs
Thanks!
Mingzhe Li, Feng Li, Mark Claypool, Bob
Kinicki WPI Computer Science Department Worcester,
Massachusetts 01609 rek_at_cs.wpi.edu

• NOSSDAV 2005
• Skamania, Washington
• June 13-14, 2005