Impact of operating points on DVFS power management
Jabran Khan, Sébastien Bilavarn, Cécile Belleudy
Abstract
Jabran Khan, Sébastien Bilavarn, Cécile Belleudy
Abstract
Dynamic voltage and frequency scaling (DVFS) techniques are commonly used for energy management in modern systems to reduce processor clock frequency and corresponding voltages. This reduces overall power and energy utilization. However the power levels associated with corresponding frequencies play an important role in effective DVFS based strategies, as in some cases energy utilization can actually be increased. We present a DVFS based power strategy dedicated to video processing with an application to an H.264 decoder. We analyze the effectiveness of this strategy using an ARM1176JZF-S hardware platform and a QEMU based virtual platform configured with models of ARM processors. Our results show that we can gain up to 57 % by choosing correct adaptation constraints. However, results also show that the efficiency of a DVFS strategy depends upon the characteristics of operating points (frequency/voltage couples). In particular, attention must be paid to the power level gap between consecutive frequencies, to compensate the increase in execution time resulting from frequency downscaling*.
OpenAlex reports 2 citations for this work. Citation counts describe recorded attention and do not establish research quality.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Dynamic voltage and frequency scaling (DVFS) techniques are commonly used for energy management in modern systems to reduce processor clock frequency and corresponding voltages. This reduces overall power and energy utilization. However the power levels associated with corresponding frequencies play an important role in effective DVFS based strategies, as in some cases energy utilization can actually be increased. We present a DVFS based power strategy dedicated to video processing with an application to an H.264 decoder. We analyze the effectiveness of this strategy using an ARM1176JZF-S hardware platform and a QEMU based virtual platform configured with models of ARM processors. Our results show that we can gain up to 57 % by choosing correct adaptation constraints. However, results also show that the efficiency of a DVFS strategy depends upon the characteristics of operating points (frequency/voltage couples). In particular, attention must be paid to the power level gap between consecutive frequencies, to compensate the increase in execution time resulting from frequency downscaling*.
Key concepts: Frequency scaling, Computer science, Power management, Power (physics), Clock rate, Voltage, Energy (signal processing), Embedded system