Modeling power consumption of high performance computer clusters

This dissertation describes a utilization-based power modeling approach that makes use of the /proc file system and is applied to estimating power load and energy consumption of high performance computer clusters. A full cluster power model is specified that meets the following conditions: 1) provides accurate energy estimates; 2) consists of a minimal number of variables; and 3) the selected variables provide a sufficient degree of explanation of the primary OS level contributors to power consumption. The cluster power model is based on analysis at the system or compute-node level, where various models are developed and evaluated. All models described in the current work are characterized by the set of variables that make up a system activity matrix (which is a sample of the utilization of a compute node from the /proc file system) and are specified by the corresponding values of the coefficients resulting from least squares minimization. System and cluster level models are evaluated using criteria that include the average of the sum of the squared residuals and the coefficient of determination. The primary assumptions in this work are 1) changes in system level utilization counters found in the /proc file system are linked to power loads captured by external Watt meters connected in series to compute nodes and 2) model coefficients derived from metered compute nodes characterize power loads non-metered compute nodes because the nodes are homogeneous. Thus, full cluster model power loads and energy consumption are obtained by applying model coefficients to cluster wide /proc based activity matrices captured for each compute node involved in a high performance computing benchmark computation. An R GUI is developed to facilitate selection of variables for the general model as well as for in-depth analysis of the links between system utilization and power consumption at the compute nodes. Results show that the /proc file system can be used to train sufficiently accurate system level models. Resulting model coefficients can be applied to activity matrices generated by arbitrary cluster nodes to estimate that node\u27s power load and total energy consumption. This approach is different from extrapolating results from a single node to all nodes of a cluster