Modeling of Aggregate Available Bandwidth in Many-to-One Data Transfer

1
Modeling of Aggregate Available Bandwidth in
Many-to-One Data Transfer

• S. C. Hui and Jack Y. B. Lee
• Department of Information Engineering
• The Chinese University of Hong Kong, Hong Kong

2
Content

• Background
• Bandwidth Availability
• Multiple-source Bandwidth Availability
• Applications
• Performance Evaluation
• Summary and Future Work

3
Background

• The Internet
• On a global scale it is still best-effort without
  end-to-end QoS guarantee.
• Presents challenges to bandwidth-sensitive
  applications (eg. Video streaming)

Client
Media Server
Playback audio/video
storage
I/O and CPU Allocation and Scheduling
I/O Resources Allocation and Scheduling
End-to-end QoS guarantee
4
Background

• Employing multiple senders has the benefits of
• Increasing the throughput
• Adapting the network bandwidth variation
• Reducing bursty packet loss
• Modeling the aggregate bandwidth

Media Server
storage
Client
Playback audio/video
Internet

5
Bandwidth Availability

• Experiments in PlanetLab 1
• Measure the bandwidth availability of 47 hosts
  using Iperf 2.

PlanetLab Hosts (sender)
PlanetLab Hosts (receiver)
s1
TCP
TCP
s2
r
. . .
Internet
TCP Throughput - averaged over 10s interval -
total 3 hours
s47
6
Bandwidth Availability

• Experiments in PlanetLab
• Average bandwidth from 0.04 to 4.53 Mbps and
• CoV from 0.16 to 0.88.

7
Bandwidth Availability

• Experiments in PlanetLab
• Bandwidth distributions

8
Bandwidth Availability

• Observations for one-to-one case
• Both mean bandwidth and CoV vary substantially
  across different senders
• Their distributions do not conform consistently

9
Multiple-source Bandwidth Availability

• Experiments in PlanetLab
• Measure the aggregate bandwidth availability of
  multiple senders.

PlanetLab Hosts (sender)
PlanetLab Hosts (receiver)
s1
TCP
TCP
s2
r
. . .
Internet
TCP Throughput - averaged over 10s interval -
total 3 hours
s47
10
Multiple-source Bandwidth Availability

• Experiments in PlanetLab
• Half of the sender-pairs have correlation
  coefficient lt 0.2
• All sender-pairs have correlation coefficients lt
  0.6.

11
Multiple-source Bandwidth Availability

• Experiments in PlanetLab
• Bandwidth distribution (sum of 47 senders)

Measured data
Normal distribution with same mean variance
12
Multiple-source Bandwidth Availability

• Experiments in PlanetLab
• Normal conformance versus number of senders
• p-value computed using the Shapiro-Wilk test 3

A p-value of 0.05 or higher is considered conform
to normal distribution.
13
Multiple-source Bandwidth Availability

• Observations for one-to-one case
• Both mean bandwidth and CoV vary substantially
  across different senders
• Their distributions do not conform consistently
• Observations for many-to-one case
• The aggregate available bandwidth of multiple
  senders tends to be normally-distributed.
• Works for as few as 4 senders (in our
  experiment).
• These are end-to-end available bandwidth
• Already accounted for the effect of link
  capacity, competing traffics, variations in the
  sender itself, dynamics of the transport protocol
  (TCP), etc.

14
Applications

• Video Content Distribution over best-effort
  networks (Internet)

Bandwidth available gtgt Video bit-rate Conventional Video Streaming
Bandwidth available ltlt Video bit-rate Hybrid Download-Streaming
Bandwidth available Video bit-rate Playback-adaptive Video Streaming
15
Applications

• Video Content Distribution over best-effort
  networks (Internet)

Bandwidth available gtgt Video bit-rate Conventional Video Streaming
Bandwidth available ltlt Video bit-rate Hybrid Download-Streaming
Bandwidth available Video bit-rate Playback-adaptive Video Streaming
16
Applications

• Video Content Distribution over best-effort
  networks (Internet)

Bandwidth available gtgt Video bit-rate Conventional Video Streaming
Bandwidth available ltlt Video bit-rate Hybrid Download-Streaming
Bandwidth available Video bit-rate Playback-adaptive Video Streaming
17
Applications

• Bandwidth available ltlt Video bit-rate
• Streaming not possible
• Download takes very long and unpredictable amount
  of time
• Playback before download possible but cannot
  guarantee playback continuity

18
Applications

• Bandwidth available ltlt Video bit-rate
• Hybrid Download-Streaming
• The goal is to shorten w and ensure playback,
  once started, will be continuous.

Partial download (w intervals)
Time
(Divides the time into intervals)
Playback begins after w intervals.
19
Hybrid Download-Streaming
Partial download (w intervals)
(Divides the time into intervals)
Time
Playback begins after w intervals.
Let the CBR movie bit rate R Let aggregate data
transfer rate for interval i Ci Total data
received by interval i Total data played back
by interval i To ensure continuous playback we
need
20
Hybrid Download-Streaming
Partial download (w intervals)
(Divides the time into intervals)
Time
Playback begins after w intervals.
Given the acceptable probability for failing
continuous playback ? To guarantee smooth
playback with the probability ?, So we need
a partial download time of
n-times autoconvolution of Cjs CDF
21
Hybrid Download-Streaming

• Trace-driven Simulation
• Bandwidth traces from PlanetLab.
• 5 10 senders, video length from 500 to 1,000
  seconds, bit-rate ranges from 200 300 kbps.
• Comparison
• Pure download download the entire video before
  playback.
• Hybrid download-streaming.
• Lower bound computed using a priori knowledge
  of future bandwidth availability (i.e., all Cis
  are known).

22
Hybrid Download-Streaming

• Simulation Results

Pure Download
Hybrid Download-Streaming
Lower Bound
23
Applications

• Video Content Distribution over best-effort
  networks (Internet)

Bandwidth available gtgt Video bit-rate Conventional Video Streaming
Bandwidth available ltlt Video bit-rate Hybrid Download-Streaming
Bandwidth available Video bit-rate Playback-adaptive Video Streaming
24
Applications

• Bandwidth available Video bit-rate
• Conventional streaming may not work well due to
  bandwidth fluctuations.
• Additional pre-buffering is needed (e.g., hybrid
  download-streaming)
• However since the available bandwidth is close to
  the video bit-rate, can we accommodate the
  bandwidth fluctuations without additional startup
  delay?

25
Applications

• Bandwidth available Video bit-rate
• Playback-adaptive Video Streaming

Video vary display frame rate Audio vary
sampling rate or use time-scale
modification
s1
s2
Playback
TCP
TCP
r
. . .
Internet
s47
Tis
Cis
If the playback rate adjustment a is small (e.g.,
a ? 5) it will not be noticeable.
26
Playback-adaptive Video Streaming

• Let the number of senders N
• Let the CBR movie bit rate R
• Let the bandwidth availability is taken at
  intervals of T sec
• The total amount of data received from all N
  senders at time interval j
• The amount of data consumed at interval j

Amount of Data received from sender i at interval
j
27
Playback-adaptive Video Streaming

• The client buffer occupancy at interval j
• By adjusting the playback rate within a small
  range, a,
• the playback time intervals can be adjusted
• Data reception rate at interval j
• Data consumption rate at interval j

28
Playback-adaptive Video Streaming

• Let Bj be the actual buffer occupancy at interval
  j.
• The estimated buffer occupancy at next interval
• The goal is to maintain buffer occupancy level,
  i.e.
• Given the client can tolerate a probability of ?
  of failing this constraint,

Possible X Prebuffer level
29
Playback-adaptive Video Streaming

• Rebuffering
• Instead using fixed rebuffer size, we calculate
  the rebuffer size as

30
Playback-adaptive Video Streaming

• Trace-driven Simulation
• 5 senders, 5 seconds initial pre-buffer.
• Video bit-rate Avg available bandwidth
  (RECi)
• Metric Avg number and duration of playback
  interruptions
• Comparison
• Conventional streaming without adaptation, pause
  and rebuffer 5 seconds video data whenever buffer
  underflows.
• Bandwidth modeling with adaptive playback, pause
  and rebuffer 5 seconds video data whenever buffer
  underflows.
• Bandwidth modeling with adaptive playback and
  adaptive rebuffering - compute the rebuffer size
  based on the estimated bandwidth model.

31
Playback-adaptive Video Streaming

• Simulation Results

Conventional streaming
Adaptive playback rebuffering
Adaptive playback only
32
Playback-adaptive Video Streaming

• Simulation Results

Conventional streaming
Adaptive playback only
Adaptive playback rebuffering
33
Playback-adaptive Video Streaming

• Simulation Results (a 5)

Number of occurrences
Duration
34
Summary and future work

• Initial Results
• Based on experiments conducted in PlanetLab
• With 4 or more senders the aggregate available
  end-to-end bandwidth is normally-distributed.
• The aggregate bandwidth availability is stable
  for many hours.
• The estimated bandwidth model can be used for
  admission control and playback scheduling to
  improve the quality of service.

35
Summary and future work

• Open Problems
• Validity of the model in the broader Internet
• How many senders are needed?
• Challenges in parameter estimation.
• How long does the measured parameter remain
  correct?
• What if some senders are correlated, e.g., pass
  through the same network bottleneck to the
  receiver?
• What if the network topology is known or
  partially known?
• Applications
• Peer-to-peer applications, distributed servers,
  etc.

36
References

• Cited Work
• PlanetLab, URL http//www.planet-lab.org/
• IPerf, URL http//dast.nlanr.net/Projects/Iperf/
• Hahn and Shapiro, Statistical Models in
  Engineering, Wiley, 1994.
• Publications
• S. C. Hui and Jack Y. B. Lee, Modeling of
  Aggregate Available Bandwidth in Many-to-One Data
  Transfer, Proc. of the Fourth International
  Conference on Intelligent Multimedia Computing
  and Networking, July 21-26, 2005, Utah, USA.
• S. C. Hui and Jack Y. B. Lee, Playback-Adaptive
  Multi-Source Video Streaming, Proc. of the
  Fourth International Conference on Intelligent
  Multimedia Computing and Networking, July 21-26,
  2005, Utah, USA.

37

• End of Presentation

38
Correlation Coefficient
The correlation coefficient is a scaled version
of the covariance between two users where