Two key technologies enabling scalable on-demand delivery of stored
multimedia content are work-ahead smoothing and multicast delivery.
Work-ahead smoothing reduces the burstiness of variable bit rate
streams, simplifying server and network resource allocation.
Recent multicast delivery techniques such as patching or bandwidth
skimming serve clients that request the same content close together
in time with (partially) shared multicasts, greatly reducing required
server bandwidth.
Although previous studies of work-ahead smoothing have generally assumed
very limited client buffer space, in a number of contexts of current
interest (e.g., systems with disked settops), it becomes feasible
to fully smooth variable bit rate content.
We quantify the start-up delay and storage requirements of full
smoothing for a number of sample video traces.
We then show that there is a fundamental conflict between aggressive
smoothing and the new multicast delivery techniques.
Work-ahead smoothing requires sending data for high rate segments of
an object earlier than it is needed for playback, while multicast
techniques yield their greatest benefit when data is delivered within
each stream as late as possible so that more clients can share that
portion of the multicast.
A new multicast delivery technique is proposed that can accommodate
aggressive smoothing with increased efficiency in comparison to
previous techniques, particularly for high request rates.
