Improving the performance of the Parallel Ice Sheet Model on a large-scale, distributed supercomputer
In this thesis, we describe our work to understand and improve the performance and scalability of the Parallel Ice Sheet Model (PISM) on the Ranger supercomputer. PISM enables the simulation of large-scale ice sheets, such as those found in Greenland and Antarctica, which are of particular interest...
| Main Author: | |
|---|---|
| Format: | Text |
| Language: | unknown |
| Published: |
DigitalCommons@UMaine
2013
|
| Subjects: | |
| Online Access: | https://digitalcommons.library.umaine.edu/etd/1895 https://digitalcommons.library.umaine.edu/context/etd/article/2924/viewcontent/MoreyTJ2013_OCR.pdf |
| Summary: | In this thesis, we describe our work to understand and improve the performance and scalability of the Parallel Ice Sheet Model (PISM) on the Ranger supercomputer. PISM enables the simulation of large-scale ice sheets, such as those found in Greenland and Antarctica, which are of particular interest to climate scientists due to their potential to contribute to sea-level rise. PISM has a unique parallel architecture that is designed to take advantage of the computational resources available on state-of-the-art supercomputers. The problem, however, is that even though PISM can run without modifcation on a supercomputer, it is generally unable to do so efficiently. We observed that PISM exhibits rapidly diminishing performance gains as the number of processors is increased, even experiencing an increase in execution time with large processor counts. PISM's inability to make efficient use of the resources available on today's supercomputers presents a challenge to researchers, particularly as larger and higher resolution data sets become available. In this work, we analyzed the reasons for PISM's poor performance and developed techniques to address these issues, resulting in an increase in performance by as much as a factor of 20. |
|---|