Measurement and characterization of Haswell power and energy consumption

Abstract

The recently introduced Intel Haswell processors implement major changes compared to their predecessors, especially with respect to power management. Haswell processors are used in the new-generation DOE NNSA tri-lab supercomputer, Trinity, hosted at Los Alamos National Laboratory. In this paper we measure and analyze a number of power-based parameter of Haswell that are of great importance for the energy consumption of applications. We study three HPC benchmarks, HPL, STREAM, FIRESTARTER and a hydrodynamics application, CLAMR. They are representative of workloads stressing different components of computers. Our experimental results show that real-time on-board power monitoring causes substantial power use if no optimization is performed; adapting P-states provides a cost-effective way to improve the power-performance of applications; enabling hyperthreading can significantly save energy by up to 96.3% for compute-bound applications; HPC applications should employ differentiated core affinity strategies in order to achieve the maximum power-performance. Moreover, we study the imbalance of sockets on a server in their power and energy use, and then propose approaches to mitigate such imbalance.