Mark Kalman
Analysis of Adaptive Media Playout for Stochastic Channel Models 

Streaming media relies on buffering at the client to counteract the
adverse effects of the Internet's best effort packet delivery service. The
client buffer smoothes out variations in packet delay, allows
retransmission attempts for lost packets, and reorders packets that arrive
out of order.  There is a trade-off, however, concerning the size of this
client buffer.  While a large buffer improves the quality of media
playback by lowering the probability that the stream is interrupted, it
also introduces a large delay between the time a media stream is requested
and the time that playback begins.

Adaptive playout schemes mitigate this tradeoff between buffer size and
delay.  The schemes control the rate of media playout to allow smaller
buffer sizes while maintaining a low probability that the buffer
underflows and the stream is interrupted.

In this project we propose to compare the performance of traditional
buffered media streams to those employing adaptive playout.  We seek to
compare the delay distributions and the probabilities of buffer underflow
for the two schemes.  We would like to show that the adaptive play-out
schemes provide better delay-underflow performance.

We will conduct the comparison first with an analysis and then with
simulations.  For the analysis, we will begin with the assumption that
packet arrivals can be modeled according to a discrete-time, switched
bernoulli process.  Using this assumption, steady-state buffer
distributions will be calculated for both the fixed and the adaptive
playout schemes.  These steady-state distributions will lead directly to
the desired values for the two schemes.  For the simulation portion of the
comparison, we will make use of software already developed by Dr.
Steinbach but modified as necessary.

In the analysis portion of the project, a major difficulty will be in the
impact of retransmissions, and modeling source rate changes in general.
Retransmissions are a complexity that are not well treated in the
literature.  Some initial thoughts on how to model the effects of
retransmissions on available bandwidth and buffer fullness include:  1)
Considering retransmissions as a further reduction in throughput during
bad periods in the channel as suggested in [3].  2)  Modeling the
channel's steady state buffer distribution first in the "good" channel
state, then in the "bad" channel state, and finally in the "recovering"
channel state.  This method would have to assume that transitions between
the "good" and "bad" are separated enough to allow a return to steady
state behavior between "bad" periods.


Work Plan:

Feb 1,  Week 1: Implement baseline buffer analysis as proposed by M.
Yuang. Prepare for proposal talk.  Consider how to handle retransmissions
and source rate changes

Feb 8,  Week 2: Explore methods to handle analysis with retransmissions
source rate changes.  Perform analysis using different approaches.

Feb 15, Week 3: Tune and finalize analytical approach.  Begin working with
Eckehard's software and simulations.  Write additions as needed.

Feb 22, Week 4: Finish simulations.  Compare with analytical results.
Explore disparities. Begin writing.

Mar. 1, Week 5: Write paper.  Prepare for talk.


References:

[1] E. Steinbach, N. Farber, and B. Girod.  Adaptive playout for low
latency video streaming. 2001

[2] Yuang, M. C., Liang, S. T. & Chen, Yu. G., "Dynamic Video Playout
Smoothing Method for Multimedia Applications", Multimedia Tools and
Applications 6, 1998, p47-60

[3] Lin et al., "Automatic-Repeat-Request Error-Control Schemes", IEEE
Comm.  Mag., Vol. 22, No. 12, pp.5-17, December 1984.

[4] M. Podolsky, S. McCanne, and M. Vetterli. Soft ARQ for layered
streaming media. Technical Report UCB/CSD-98-1024, University of
California, Computer Science Division, Berkeley, CA, November 1998.

[5] Robertazzi, Thomas G., Computer Networks and Systems - Queuing Theory
and Performance Evaluation, second edition, Springer Verlag, 1994