Dynamic Control of Scalable Media Processing Applications

Abstract

Media processing in software is characterized by highly fluctuating, content-dependent processing times, as well as strict real-time requirements. These realtime requirements can be met by means of a worst-case resource allocation, but this is often not cost-effective. To assign resources closer to the average-case load situation, scalable media processing may be applied. A scalable media processing application allows a trade-off between the resource usage and the output quality, by varying the quality level at which processing is done. We consider the problem of controlling the quality level of a scalable media processing application during run time. The objective is to maximize the quality of the output as perceived by the user, by simultaneously maximizing the quality level and minimizing the number of artifacts that result from both deadline misses and quality-level changes.

First we model the problem as a finite Markov decision process. We use this to determine on-line quality-level control strategies. These strategies enable us to cope with temporal load fluctuations. Next, to handle structural load fluctuations, we present a method called scaled budget control. The scaled budget control strategy works well, as we show by several simulation experiments using a scalable MPEG-2 decoder, where its performance comes close to an off-line computed upper bound.