On Packet Loss Prediction for Realtime Packet Audio

1
On Packet Loss Prediction for Real-time Packet
Audio

• Lopamudra Roychoudhuri and Ehab Al-Shaer
• DePaul University
• School of Computer Science
• Chicago, IL
• lroychou,ehab_at_cs.depaul.edu

2
Presentation Outline

• Background and Motivation
• Related Work
• Objective
• Approach
• Pending Work
• Conclusion
• QA

3
Background and Motivation

• Quality of an audio communication is highly
  sensitive to packet loss
• UDP/IP, used for audio communications, does not
  provide packet retransmission
• Also, packet-retransmission is not a viable
  option for real-time audio, since it adds to the
  delay that might exceed allowable mouth-to-ear
  delay

4
Background and Motivation cont.

• Packet loss for audio can be rectified by adding
  redundancy using Forward Error Correction (FEC)
• But FEC has bandwidth and processing overhead
• It will be immensely helpful to an audio
  transmission if we can add FEC only when it is
  necessary

5
Background and Motivation cont.

• In the Internet experiments we have observed that
  there seems to be some correlation between delay
  increase and packet loss

Fig.2 Korea 12/18/02 84920am
6
Related Work

• Vern Paxson (Ph.D. Thesis)
• he concluded that the linkage between delay
  variation and loss was weak, though not
  negligible
• Sue B. Moon (Ph.D. Thesis)
• ran experiments on the Internet
• measured per-packet delay and packet loss
• reported a quantitative study of correlation
  patterns
• but they did not attempt to predict packet loss
  from the delay variation data of an ongoing
  communication

7
Related Work cont.

• Jain Dovrolis
• In a tool called Pathload they use the delay
  variation principle to measure available
  bandwidth
• They send streams of increasing rate till the
  stream saturates the available bandwidth and
  starts showing distinct increase in delay
• Packet loss is highly probable when the available
  bandwidth is really low and is consumed by the
  ongoing cross traffic
• Our research methods and experiments are based on
  this premise

8
Objective

• The basic idea behind an ongoing loss prediction
  method is to successfully track the increase and
  decrease in One Way Delay (OWD)
• Accordingly predict packet loss with high
  probability from observing the changes in delay
  variations due to congestion in the link leading
  to lack of available bandwidth
• The task becomes difficult due to the
  unpredictable and dynamic nature of cross traffic
  in the Internet

9
Technical Approach

• Internet Experiments
• Description
• experiments on the RON/NetBed from emulab.net
• A program (Master) from the RON site at MIT,
  MA, USA
• sent a speech segment of 1 minute to various Unix
  RON sites across the world (Korea, Netherlands,
  Lulea and Greece)
• Observations
• Baseline delay
• Steady-state delay, signifying the delay under no
  congestion
• Loss Threshold delay
• Delay at capacity saturation point

10
Observations cont.
Fig. 3. Korea 12/18/02 84920 Baseline Loss
Threshold
11
Technical Approach - Simulation

• Simulation Experiments on ns2
• CBR stream is transmitted over a path consisting
  of four nodes of a certain capacity
• An interim Pareto cross-traffic uses part of the
  path and saturates the available bandwidth for
  time being

Cap 1.6m,CBR 1.2m,pareto 550k
Cap 1.6m, CBR 1.2m, pareto 800k
12
Loss Predictor Metrics

• Delay Distance Metric
• gives an absolute ratio of the delay value in
  relation to the threshold
• It indicates how much the delay has increased
  over the baseline
• Delay Trend Metrics
• The trend metrics indicate how fast the delay is
  increasing

13
Delay Trend Metrics

• Short Term Metrics
• give indications of upward or downward trends for
  short-term (previous 5 to 10 packets)
• Long Term Metrics
• give indications of upward or downward trends for
  long-term (previous 20 to 50 packets)
• By considering the rate of increase, the
  short-term and the long-term metrics prevent
  Delay Distance metric from over-reacting to the
  absolute value of the delay

14
Delay Distance Metric

• min_max min1 , (Dk bl)/(thr bl)
• where,
• bl the most observed delay so far, considered
  to be the baseline delay
• Dk the delay value of the k-th packet,
• thr the threshold delay at which a loss is
  observed

thr
Loss threshold delay
Dk
Baseline delay
bl
bl
15
Short Term Metrics

• tan-slope
• determines how fast the delay is approaching the
  loss threshold
• indicator of sharpness of increase (the slope
  of the increase)
• tan ? max-1,min1,(Dk Dk-1)/(tk - tk-1)

16
Short Term Trend cont.

• SPCT/SPDT
• indicators of consistent increasing trends
• how consistently the delay is increasing for
  every measured packet pair

Where I(X) is 1 if X holds, and 0 otherwise
17
Long Term Metrics

• SPCT_lt/SPDT_lt long term versions of SPCT and
  SPDT (20 to 50 packets)
• min_max_avg_lt Averages min_max for last 20 to
  50 packets

18
Formalization of Loss Predictor Algorithm

• predictor 0 f(min_max, short_term_trend,
  long_term_trend) 1
• where,
• f(min_max, short_term_trend, long_term_trend)
• w1 min_max w2 short_term_trend w3
  long_term_trend
• min_max fa(delay distance, loss threshold)
  min1 , (Dk bl)/(thr bl)
• short_term_trend fb(SPDT, tan-slope) (SPDT
  tan-slope) / 2
• long_term_trend fc(min_max_avg_lt, SPDT_lt,
  SPCT_lt)
• (min_max_avg_lt SPCT_lt SPDT_lt) / 3
• w1 w2 w3 1

19
Zones based on min-max value
20
Weight rules

• w1 w2 , w3
• w2 0, w3 increasing in Yellow zone, when
  0 min_max lt 65
• w2 increasing, w3 decreasing in Orange zone,
  when 65 min_max lt 75
• w2 increasing, w3 0 in Red zone, when 75
  min_max 100

21
Weight Function Prototype Set - sqrt

• Predictor(f1min_max f2LT f3ST)
• where
• f1 1 sqrt(min_max)/2,
• f2 sqrt(min_max)/2f4
• f3 sqrt(min_max)/2(1-f4)
• f4 1 in Yellow
  zone, 0 min_max lt .65
• 1-(min_max-.65)10 in Orange zone, .65
  min_max lt .75
• 0 in Red
  zone, .75 min_max 1

22
Weight Function Set - sqrt
23
Preliminary Results
Fig. 8 Korea 12/18/02 84920am, Effect of
Short-term and Long-term trend on Predictor
24
Pending Work

• Refine the metrics and the weight factors for
  better predictor efficiency and accuracy
• More experimentation and analysis will give us
  further insight into the patterns of delay-loss
  correlation
• Extend the concepts of a loss predictor for
  Random Early Detection (RED) packet dropping
  mechanisms and for multicast communication

25
Conclusion

• we present a novel mechanism to predict packet
  loss by observing the changes in the available
  bandwidth in terms of short-term and long-term
  variations in delay
• Loss Predictor approach is based on
• Determining certain metrics from the ongoing
  traffic delay variation characteristics,
• Combining them with different weights based on
  their importance, and
• Deriving a predictor value as the measure of the
  packet loss probability
• We present some preliminary results showing the
  efficiency and the accuracy of the algorithm

26
QA