The reservation rate problem that arises in a distributed audio/video editing system is to estimate and then reserve the resources--including shared remote disks, SCSI buses, and FDDI network capacity--required to play a document consisting of audio and video segments. Ideally, one wants to maximize resource usage without dropouts, where a dropout occurs when an audio or video segment does not arrive in time to play.
The author describes and compares three algorithms for estimating the fraction of the capacity of each device required to play the document and then requesting a reservation of the fraction. The first algorithm, based on long-term average device usage by segments, is fast but imprecise. The second algorithm produces more accurate rate estimates but requires much higher overhead. The third algorithm, a hybrid of the first two, uses the long-term averages as estimates when the set of active segments is static and uses the more accurate estimates when the set of active segments changes. In addition to comparing the three reservation rate algorithms, the author discusses the modeling and measurement of the performance of the critical resources.
Examples supported by figures illustrate each of the algorithms. This clear, well-organized paper provides a practical example of real-time resource allocation in a distributed system.