Auto-Tuning Parallel I/O

Investigators: Suren Byna and Prabhat (LBNL), Quincey Koziol (The HDF Group), Venkat Vishwanath (ANL)

Problem

Obtaining good I/O performance on modern HPC systems is challenging due to complex hardware and software dependencies.

Approach

We are developing a auto-tuning system for obtaining a significant fraction of peak I/O performance.

Accomplishments

• Developed a genetic algorithm and statistical framework for efficiently searching parameter space. Published results at SC'13 and HPDC'14.
• Tested framework on 3 scientific benchmarks on Cray, IBM and Sun platforms
• Demonstrated 2-50X improvement in write performance over default system settings

Impact

• Enables science codes to obtain high I/O resource utilization without code modification