measurements of congestion responsiveness of windows streaming media (wsm)
DESCRIPTION
Measurements of Congestion Responsiveness of Windows Streaming Media (WSM). Presented By:- Ashish Gupta. Roadmap. Introduction Graphs and Observations Conclusion Comments. Introduction. AIM: Characterize the bitrate response of “Intelligent streaming” and content “encoding” of WSM - PowerPoint PPT PresentationTRANSCRIPT
Measurements of Congestion Responsiveness of Windows Streaming Media (WSM)
Presented By:-
Ashish Gupta
Introduction
AIM: Characterize the bitrate response of
“Intelligent streaming” and content “encoding” of WSM
Variable Parameters: Network characterstics like bottleneck
bandwidth Content encoding characterstics
Windows Streaming Media (WSM)
Intelligent Streaming: Content bitrate is adjusted according to current
available bandwidth Server and Client determine current bandwidth
Multiple bitrate encoded files are used for streaming
Intelligent degradation in image quality is done first (both server and client support)
Audio is reconstructed to preserve quality
WSM (Contd.)
WSM server decides the current available bandwidth
It requires multiple bitrate encoded stream support by the media creator
Some media data parameters like Thinning, Maximum bitrate etc. can be set during transfer initiation.
Setup (Contd.)
Client has media tracker s/w for capturing the client side performance statistics
Each side of the router has linux machine capture offered and achieved network loads.
Internet behavior is achieved by router by setting the internet latencies (normal internet traffic)
Graphs
All graphs in the next few slides consider single bitrate encoding
Network bandwidth 725 Kbps and latency set to measured internet latency equal to 45 ms
WSM provided with 60 sec clip encoded at 340 Kbps, roughly equal to fair capacity
Observations
20-40 percent of packet loss occurred during buffering phase
WSM playing can be partitioned into buffering phase and playout phase
Previous experiment is repeated with 540Kbps clip which is much above the fare share bandwidth
Observations
Media Thinning functionality provided by “intelligent streaming” was NOT used
Competing TCP flow was denied of fair share of available capacity
During buffering period heavy packet losses were encountered
Observations
No media thinning during the buffering phase
Bitrate is increased even during losses during buffering phase (Fire-hose approach)
Transmission rate decreased below fair share bandwidth during playout phase
Graph
Now buffering phase and post-buffering phase is studied independently
Graphs are plotted “transmission bitrate of WSM/TCP” vs
“Content encoded bitrate” “loss rate” vs “Content encoded bitrate”
Different capacity n/w is considered without induced losses for “Buffering phase only”
Observations
Buffering rate is proportional to content encoding rate until the encoding rate exceeds bottleneck capacity
Beyond this loss rate can be as high as 80%
WSM has high loss rate due to higher sending rate
Observation (Contd.)
The behavior of WSM changes between 340 Kbps-548 Kbps
At 548Kbps encoding rate is there is a significant drop in loss rate
Observations
In post-buffering bit rate is proportional encoding bit-rate till bottleneck capacity
For still higher encoding bit rate incurs losses upto 40 percent initially
Bitrate goes lower then TCP flow due to media thinning
Accurate information of “thinning” behavior over post buffering period is missed due to average values
Observations
WSM traffic is bursty in nature
Retransmission happens for any dropped packet
Reason for spikes: shortening of transmission time between packets bursts
Graphs (Multiple bit rate)
Aim: Explore number of bitrate vs responsiveness
Ten set of clips were used 1st clip (1128 Kbps) 2nd clip (1128, 764 Kbps) . 10th clip (1128,.. 764, 548,.., 282, 148..Kbps)
Study of Bitrate vs lowest bitrate contained in the clip is made
Graph (Buffering: Multiple bit rate)
For Buffering: WSM chooses the bitrate which is just lower than bottleneck bandwidth otherwise it takes the lowest capacity available
WSM specific observations
Bottleneck bandwidth: Available bandwidth is measured in WSM using
three large packets. Two packet pair estimate is used to get the bottleneck bandwidth
Bottleneck bandwidth is estimated only once before each session and then RTCP messages are used to control retransmission
Frequency of RTCP message increases with increasing loss rate
Graphs (Induced losses)
Loss induced by router to check the behavior of WSM in case of loss due to network
For further discussion bottleneck bandwidth: 725 Kbps Encoded bitrate: 548 Kbps (largest below 725
Kbps)
Observations (Buffering: network induced loss) With high loss induced loss rate
TCP decreases its flow WSM increases its flow to compensate for
high loss rate
During the initial losses WSM buffering uses “fire-hose” approach
Observations (post-buffering: network induced loss) Loss rate of 3-5% causes WSM to thin stream
Above 5% loss rate there is no change in thinned bitrate
Graphs (Induced losses MBR clip)
Next experiment uses multiple bitrate (MBR) encoded ( 548, 340, 282, 148, 106, 58…. Kbps)
Graphs (sudden change in induced loss)
This shows that WSM is unaffected by the sudden change in network loss
WSM model
WSM model: Should have two phases buffering and post
buffering Bursty Buffering is TCP unfriendly In some post buffering period WSM is more
than TCP friendly
Conclusion
During buffering TCP friendliness can only be achieved if encoded bitrate is less then estimated capacity. Otherwise, buffering is done at encoding rate
During playout Thinning or streaming low bitrate encoded
stream is done This can also be TCP unfriendly if encoding
rate is more than half capacity and less than full capacity
Conclusion
This study shows that content provider can make judicious decision encoding rates number of encoding levels
It provides the researcher with more accurate model of streaming media traffic