| Summary: | Most of the existing analyses on admission control and buffer management
for continuous media streams assume fixed rate data compression. This assumption
is invalid for MPEG streams, which have variable bit rate (VBR) compression. With
the increased acceptance of VBR compression and the development of new, more
efficient VBR techniques, there is a need for new models, analyses and algorithms
for streams with variable compression. The Central Limit Theorem of Statistics has
been proposed for use at the granularity level of data streams to handle admission
control of concurrent variably compressed streams. The implication here is that the
total amount of buffer space required by all streams in the system at a particular
point in time is approximately normally distributed. In this thesis we develop
a model for MPEG streams that applies the Central Limit Theorem at the finer
granularity level of frames. This gives a stronger and more general result: that the
amount of buffer space required for each stream approximates a normal distribution.
Using this model we develop several admission control algorithms that provide
user-selectable, individual non-overflow guarantee levels by computing the amount
of exclusive buffers needed to provide these guarantees. Experimental results indicate
that the buffer space overhead required to support individual guarantees is fairly small and worthwhile. We also investigate the feasibility of providing an additional
q% system-wide non-overflow guarantee on top of the individual guarantees
through the use of shared buffers. Experimental results indicate that the buffer
space overhead is again quite small but very successful in enhancing the reliability
and quality of service to the user.
|
|---|
