Impact of power-management granularity on the energy-quality trade-off for soft and hard real-time applications

In this paper we introduce the concepts of work of tokens (e.g. video frames) in an application, and slack arising from variations in work. Slack is used for dynamic voltage and frequency scaling in combination with a conservative power-management policy that never misses deadlines, for hard real-time applications, and with a non-conservative policy for soft real-time applications. We evaluate both policies for a number of different granularities (frequency of activation of the power manager) on an MPEG4 application, on energy and quality (deadline misses). We conclude that for soft real-time applications, there is a clear optimum in the energy, which depends on the work histogram of the application. The conservative policy has no deadline misses, and is only negligibly more expensive in terms of energy than the non-conservative policy. Finally, the granularity of both policies can be very coarse (128 frames) to reduce the power manager activation frequency, which has an insignificant energy cost.