Evaluating Parallel I/O Energy Efficiency

Energy efficiency and parallel I/O performance are two primary constraints in high performance computing (HPC). Large scale computing systems consume vast amounts of energy and require scalable I/O performance to meet the needs of a broad spectrum of data intensive scientific applications. However, little research has provided an in-depth understanding of energy efficiency in parallel I/O subsystems. In this paper, we experimentally investigate the efficiency impact of a wide range of variables including the application's access patterns, parallel file system deployment, and processor frequency scaling. Our studies lead to several key findings on parallel I/O energy efficiency. First, the application's I/O access pattern significantly affects system energy efficiency. By aggregating file accesses and choosing appropriate data sizes, we can both improve performance and reduce energy use. Second, parallel file systems, such as PVFS, not only increase the I/O scalability but also are more energy efficient. Matching the application's I/O buffer size to the file system strip size will lead to much higher efficiency than poorly chosen I/O buffer sizes. Third, DVFS can further reduce energy use of parallel I/O operations with little performance impact. These findings are useful to guide the development of energy efficient HPC systems and software.