Adaptive Media Streaming - Scalable Continuous Media Streaming Systems - page 148

B T , then it implies that the client buffer occupancy is below the thresh-
old. Hence the server will reduce the video bit-rate to raise the buffer occupancy to B T by
substituting B n n + 1 =

Now if B n i <

B T in equation (8.9) to obtain:

1 D i + 1

B T − B n i

r i + 1 = − (8.10)

M

Otherwise if B n i B T , then it implies that the client buffer occupancy is above the threshold.

In this case the server will simply maintain the current client buffer occupancy by setting
B n i + 1 =

≥

B n i in equation (8.9) to obtain r i + 1 . This is a conservative strategy to reduce the
likelihood of buffer underflow. Thus, we have:

D i + 1

r i + 1 = (8.11)

Finally, the server checks and limits the computed video bit-rate to the feasible range
[ r min , r max ]by

r i + 1 = min { r max , max { r min , r i + 1 }} (8.12)

Note that in contrast to previous works [1-3], this adaptation algorithm has no control
parameter that requires either offline or online optimization. This has practical significance as
it is not easy to optimize the control parameters without knowledge of the available network
bandwidth.

8.5.2 Preemptive Rate Control
In our trace-driven simulations, we find that the available network bandwidth can occasionally
drop drastically to a very low value. These sudden bandwidth drops do not appear to be
predictable and thus can result in client video playback starvation.
The fundamental problem is that the adaptation algorithm is executed only when a new
video segment is to be transmitted. Thus, if bandwidth drops significantly, then the trans-
mission of the current video segment will stall. The adaptation algorithm cannot react in
this case as the current video segment has not yet been completely transmitted. Meanwhile
the client will continue consuming video data for playback and thus may run into buffer
underflow.
To tackle this problem, we can use a preemptive scheduling technique to shorten the delay for
the adaptation algorithm to react to changing network conditions. Instead of waiting indefinitely
for a video segment to be completely submitted into the server buffer, the scheduler will timeout
after Mr i + 1 / D i + 1 seconds, which is the expected time required to submit the ( i

1)th video
segment into the server buffer. If not all video data can be submitted, then the remaining
yet-to-be-submitted data will be discarded and the remaining video segment transcoded again
according to the new estimates on the client buffer occupancy and the available network
bandwidth.
Note that preemptive rate control requires the video transcoder to be able to adjust the video
bit-rate in between a video segment. The implementation will be highly dependent on the video
compression algorithm employed and further study is required to identify the constraints and
tradeoffs of this technique.

+