Automatic MPI-IO Tuning with the Periscope Tuning Framework
Weifeng Liu, Isaías A. Comprés Ureña, Michael Gerndt, Bin Gong · 2014
The performance of I/O operations on HPC systems is a bottleneck in many parallel applications. The MPI Forum defined the MPI-IO programming interface for parallel I/O as part of the MPI-2 standard. With MPI-IO, parallel applications can overcome the performance and portability limitations of existing parallel I/O interfaces. MPI-IO performance analysis and tuning are important for parallel software development. Some analysis tools have been designed to help developers understand MPI-IO performance. However, there are no recommendations given by those tools on how to tune the MPI-IO code to obtain better performance. In this paper, we introduce the design of an MPI-IO automatic tuner. The automatic tuner relies on the Periscope Tuning Framework (PTF) for performance data collection and for passing hints to the MPI-IO library. The goal is to produce tuning recommendations that will help applications achieve efficient execution of MPI-IO operations in production runs.