SIParCS 2015- Zhang

Qian Zhang, Louisiana State University

Evaluating PIO2 performance in the CESM-POP Model

(Slides) (Recorded Talk)

The Parallel Ocean Program (POP) is one of the most computationally intensive components in the CESM model.   Proportionally POP has one of the most expensive IO costs in CESM therefore optimizing the I/O time can significantly  improve POP performance.  The Parallel I/O (PIO) library enables multiple processes to participate in reading data from or writing numerical simulation data into a common file in parallel. Recently a new version of the PIO library (PIO2) has been introduced which has the potential to significantly improve IO performance.  In this study we evaluate PIO2 performance in the POP model and look for opportunities to improve it. 

By comparing with PIO1 (baseline case), under different data rearrangement algorithms and various scalability conditions as well as  adding timers and using profiling analysis (gprof), the hotspots can be identified for future optimization.

Mentor: Rory Kelly, CISL