K. Iskra, K. Yoshii, R. Gupta, P. Beckman, "Power Management for Exascale," Preprint ANL/MCS-P3011-0712, July 2012. [pdf]
In order to meet the exascale goals, today's top HPC systems will need to scale by two orders of magnitude, at the same time increasing their power consumption by only an integer factor. Power consumption of both individual nodes and the overall system is thus a critical issue to address .
Most performance studies of large-scale HPC systems and their workloads have focused primarily on flops, bandwidth, and latency. Few concrete studies exist that focus on quantifying power and energy consumption at the hardware and software levels. Until recently, system vendors have had little incentive to expose extensive system and component-level power interfaces to users. Consequently, the power-management methodology is lacking, and the underlying capabilities in today's computer systems are limited or missing. Exascale systems, consisting of hundreds of thousands of nodes drawing tens of megawatts of electrical power, mandate a need for new, systemwide methodologies and procedures for power monitoring, management, and scheduling.