Low power is an important design requirement for HPC systems nowadays. Dynamic voltage and frequency scaling (DVFS) has become the commonly used and efficient technology to achieve a trade-off between power consumption and system performance. However, most of the prior work using DVFS did not take into account the latency of voltage/frequency scaling, which is a critical factor in real hardware determining the efficiency of the power management algorithm. This paper investigates the latency aspects of DVFS on a real many-core hardware platform. We propose a latency-aware DVFS algorithm to achieve profile-guided power management to avoid aggressive power state transitions. We evaluate our algorithm on the Intel SCC platform using a data-intensive benchmark, Graph 500. The experimental results not only show impressive potential for energy saving in data-intensive applications (up to 31% energy saving and 60% EDP reduction), but also evaluate the efficiency of our latency-aware DVFS algorithm which achieves 12.0% extra energy saving and 5.0% extra EDP reduction, while increasing the execution performance by 22.4%.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.