Reducing Idle Power Consumption in High Performance Systems

Although high-performance computing has always been about efficient application execution, both energy and power consumption have become critical concerns owing to their effect on operating costs and failure rates of large-scale computing platforms. Modern processors provide techniques such as Dynamic Voltage and Frequency Scaling (DVFS) and Clock Modulation (Throttling), which modify their performance and power consumption to reduce the dynamic power consumption. Furthermore, with the continual reduction in semiconductor feature size, the idle power consumption will be dominating the dynamic power consumption in the foreseeable future. In this work, a runtime monitoring method is proposed to reduce the idle power consumption of compute nodes in a cluster based on their current utilization. Experiments performed on the clusters tested depict that the proposed method enable energy savings of up to 87%.