Energy Saving Potential of Idle Pacman Supercomputing Nodes

Abstract

To determine the energy saving potential of suspending idle supercomputing nodes without sacrificing efficiency, my research involved the setup of a compute node power usage monitoring system. This system measures how much power each node draws at its diff erent levels of operation using an automated Expect script. The script automates tasks with interactive command line interfaces, to perform the power measurement readings. Steps required for the power usage monitoring system include remotely logging into the Pacman Penguin compute cluster power distribution units (PDUs), feeding commands to the PDUs, and storing the returned data. Using a Python script the data is then parsed into a more coherent format and written to a common file format for analysis. With this system, the Arctic Region Supercomputing Center (ARSC) will be able to determine how much energy is used during diff erent levels of load intensity on the Pacman supercomputer and how much energy can be saved by suspending unnecessary nodes during levels of reduced activity. Power utilization by supercomputers is of major interest to those who design and purchase them. Since 2008, the leading source of worldwide supercomputer speed rankings has also included power consumption and power efficiency values. Because digital computers utilize electricity to perform computation, larger computers tend to utilize more energy and produce more heat. Pacman, an acronym for Pacific Area Climate Monitoring and Analysis Network, is a high performance supercomputer designed for large compute and memory intensive jobs. Pacman is composed of the following general computational nodes: • 256 four-core compute nodes containing two dual core 2.6 GHz AMD Opteron processors each • 20 twelve-core compute nodes containing two six core 2.6 GHz AMD Opteron processors each • 88 sixteen-core compute nodes containing two eight core 2.3 GHz AMD Opteron processors each