1 Measuring Congestion Responsiveness of Windows Streaming Media James Nichols Advisors: Prof. Mark Claypool Prof. Bob Kinicki Reader: Prof. David Finkel.

1

Measuring Congestion Responsiveness of Windows

Streaming Media

James NicholsAdvisors:

Prof. Mark ClaypoolProf. Bob Kinicki

Reader:Prof. David Finkel

Thesis PresentationPEDS - 12/8/03

2

Network Impact of Streaming MediaNetwork Impact of Streaming Media

• Unlike file transfer or Web browsing, Streaming Media has specific bitrate and timing requirements.

• Typically, UDP is the default network transport protocol for delivering Streaming Media.

• UDP does not have any end-to-end congestion control mechanisms.

3

The Dangers of UnresponsivenessThe Dangers of Unresponsiveness

• Flows in the network which are unresponsive to congestion can cause several undesirable situations:• Unfairness when competing with responsive flows

for limited resources• Unresponsive flows can contribute to congestion

collapse

• Some Streaming Media applications use UDP, but rely on the application layer to provide adaptability to available capacity

• Performance of these application layer mechanisms is unknown

4

Intelligent StreamingIntelligent Streaming

• Application layer mechanism of Windows Streaming Media (WSM) to adapt to network conditions

• Can “thin” streams by sending fewer frames• If the content producer has encoded multiple

bitrates into the stream, IS can choose an appropriate one

• Chung et al. suggests that technologies like IS may provide responsiveness to congestion, even TCP-friendliness

• Performance of Intelligent Streaming is unknown

5

Research GoalsResearch Goals

• No measurement studies have been completed where researchers had total control over:• The streaming server• Content-encoding parameters• Network conditions at or close to the server

• We seek to characterize the bitrate response function of Windows Streaming Media in response to congestion in the network.

• Want to precisely quantify relationship between content encoding rate and performance.

6

OutlineOutline

• Introduction• Related Work • Methodology• Results & Analysis• Conclusions• Future Work

7

Related WorkRelated Work

• Some research has been done in the general area of Streaming Media:• Traffic characterization studies [VAM+02, dMSK02]

performed through log analysis• Empirical studies using custom tools [CCZ03,

WCZ01, LCK02]

• Characterization of streaming content available on the Web [MediaCrawler]

• None had control of the server, client, and network conditions

8

Need control over the serverNeed control over the server

• Not having server limits possible data set of content to study

• For example, [LCK02], measured IP packet fragmentation when streaming WSM clips but packet size can be tuned server-side

• Other research [CCZ03] had to stream over the public Internet while measuring network performance

9

OutlineOutline

• Introduction• Related Work• Methodology • Results & Analysis• Conclusions• Future Work

10

MethodologyMethodology

•Construct testbed

•Create/adapt tools

•Encode content

•Systematic control

•Examine SBR clip

•Range of SBR clips

•MBR clips

•Vary loss and latency

11

Results and Analysis

Single Bitrate Clip

12

ExperimentsExperiments

• Single bitrate (SBR) clip in detail • Range of SBR clips• Multiple bitrate (MBR) clips• Additional experiments performed but

not discussed here

13

Single Bitrate Clip ExperimentSingle Bitrate Clip Experiment

• Hypothesis: SBR clips are unresponsive to congestion

• Latency: 45 ms• Induced loss: 0%• Bottleneck capacity: 725 Kbps• Start a TCP flow through the link• 10 Seconds later stream a WSM clip• Measure achieved bitrates and loss rates

for each flow

14

340 Kbps Clip - Bottleneck Capacity 725 Kbps340 Kbps Clip - Bottleneck Capacity 725 Kbps

< 0.001 packet lossAfter 15 seconds

TCP-Friendly?

15

548 Kbps Clip - Bottleneck Capacity 725 Kbps548 Kbps Clip - Bottleneck Capacity 725 Kbps

~ 0.003 packet loss for WSM~ 0.006 packet loss for TCP

after 15 seconds

NotTCP-

Friendly!

16

1128 Kbps Clip - Bottleneck Capacity 725 Kbps1128 Kbps Clip - Bottleneck Capacity 725 Kbps

Responsive!

17

Network TopologyNetwork Topology

18

Measuring Buffering PerformanceMeasuring Buffering Performance

• Parse packet capture for RTSP PLAY message

• Examine MediaTracker output and measure how long it took from the start of streaming to when the buffer is reported to be full

• PLAY + interval = buffering period

19

ExperimentsExperiments

• SBR clip in detail• Range of SBR clips • MBR clips

20

Comparison of Single Bitrate ClipsComparison of Single Bitrate Clips

• Want to precisely quantify relationship between content encoding rate and performance

• Repeat the previous experiment, but for a range of single bitrate clips:• 28, 43, 58, 109, 148, 282, 340, 548, 764, 1128 Kbps

• Vary network capacity: 250, 750, 1500 Kbps

• Measure performance during and after buffering

21

SBR Clips - Bottleneck Capacity 725 KbpsBuffering Period

SBR Clips - Bottleneck Capacity 725 KbpsBuffering Period

22

SBR Clips - Bottleneck Capacity 725 KbpsPlayout Period

SBR Clips - Bottleneck Capacity 725 KbpsPlayout Period

23

Results and Analysis

Multiple Bitrate Clips

24

Multiple Bitrate ClipsMultiple Bitrate Clips

• Hypothesis: Multiple Bitrates make WSM more responsive to congestion

• Same experiment as before, but with different encoded content

• Vary network capacity: 250, 725, 1500 Kbps

• Created two sets of 10 multiple bitrate clips• Experiments with lots of other MBR clips

25

Multiple Bitrate ContentMultiple Bitrate Content

• First set of clips (adding lower):• 1128 Kbps• 1128-764 Kbps• 1128-764-548 Kbps• …• 1128-764-548-340-

282-148-109-58-43-28 Kbps

• Second set of clips (adding higher):• 28 Kbps• 28-43 Kbps• 28-43-56 Kbps• …• 28-43-58-109-148-

282-340-548-764-1128 Kbps

26

Adding lower bitrates to clip - 250 Kbps Bottleneck Capacity - Buffering PeriodAdding lower bitrates to clip - 250 Kbps Bottleneck Capacity - Buffering Period

27

Adding lower bitrates to clip - 250 Kbps Bottleneck Capacity - Playout Period

Adding lower bitrates to clip - 250 Kbps Bottleneck Capacity - Playout Period

28

Adding lower bitrates to clip - 725 Kbps Bottleneck Capacity

Adding lower bitrates to clip - 725 Kbps Bottleneck Capacity

Buffering Playout

29

Adding higher bitrates to clip - 725 Kbps Bottleneck CapacityAdding higher bitrates to clip - 725 Kbps Bottleneck Capacity

Buffering Playout

30

Additional experimentsAdditional experiments

• Not enough time to discuss all the results• Different bottleneck capacities• Carefully choose 2 or 3 bitrates to include

in MBR clips• Vary loss rate• Vary latencies• Also look at other network level metrics:

interarrival times, burst lengths, and IP fragmentation

31

ConclusionsConclusions

• Prominent buffering period means WSM cannot be modeled as a simple CBR flow

• WSM single bitrate clips:• During buffering WSM responds to capacity

only when the encoding rate is less than capacity

• Otherwise, high loss rates are induced

• During playout WSM responds to available capacity

• Thin if necessary• If rate is less then capacity, will still be responsive to

high loss rates (5%)

32

ConclusionsConclusions

• WSM multiple bitrate clips:• During buffering WSM responds to capacity

only when content contains a suitable bitrate to choose

• Chosen bitrate is largest that capacity allows• Otherwise, still tries to fit the smallest available,

again resulting in high amounts of loss

• During playout WSM is responsive to available capacity

• Either because it chose the proper rate, or because it thins if proper rate isn’t encoded in clip

• However, the chosen bitrate probably isn’t fair to TCP

33

Future WorkFuture Work

• Run the same experiments with other streaming technologies: RealVideo and Quicktime

• Examine the effects of different content types

• Build NS simulation model of streaming media for use in future research

34

QuestionsQuestions