Analyzing Energy-Time Tradeoff in Power Overprovisioned HPC Data Centers

Abstract

Minimizing energy and power consumption of large scale data centers is one of the biggest challenges faced by the high performance computing community. In an over provisioned data center, nodes are power capped to run below their Thermal Design Power (TDP) value and therefore, an over provisioned data center has more nodes than a conventional data center with the same power budget. In this work, we study the energy versus time trade-off in a power over provisioned HPC data center. We show that over provisioning with the goal of maximizing performance can lead to excessive energy consumption. However, careful selection of configuration, that is number of nodes and power cap, can lead to significant savings in energy consumption with very small penalty on execution time of the application. We achieve up to 15% savings in energy consumption with only 2.8% increase in execution time as compared to a configuration that yields the best execution time.