Assessing Energy Efficiency of Parallel I/O Systems
SESSION: Research Poster Reception
EVENT TYPE: Poster
TIME: 5:15PM - 7:00PM
AUTHOR(S):Rong Ge, Xizhou Feng, Jiping Hu, Xian-He Sun
ABSTRACT: Energy efficiency and parallel I/O performance are two primary constraints in today’s high performance computing systems design. However, little research has provided an in-depth understanding of energy efficiency in parallel I/O subsystems. We experimentally investigate the energy efficiency of parallel I/O systems on a DVFS (Dynamic Voltage Frequency Scaling) enabled power-aware multicore cluster. We evaluate the performance, energy, and efficiency of parallel I/O with NFS, Lustre and PVFS, and study how the efficiency is impacted by a wide range of variables including the application’s I/O access patterns, parallel file system deployments, and processor speed scaling. We found: the application’s I/O access pattern significantly affects system energy efficiency; Tailoring parallel file system design and deployment to match the application's I/O characterisitics is critical for energy efficiency; DVFS can reduce energy consumption of parallel I/O operations with little performance impact. These insights will provide guidelines for developing energy efficient HPC systems and software.