Perfromance analysis of the Parallel Community Atmosphere Model (CAM) application
Efficient execution of parallel applications requires insight into how the parallel system features impact the performance of the application. Significant experimental analysis and the development of performance models enhance the understanding of such an impact. Deep understanding of an application...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | en_US |
| Published: |
2010
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/1969.1/ETD-TAMU-1761 http://hdl.handle.net/1969.1/ETD-TAMU-1761 |
| Summary: | Efficient execution of parallel applications requires insight into how the parallel
system features impact the performance of the application. Significant experimental
analysis and the development of performance models enhance the understanding of such
an impact. Deep understanding of an application’s major kernels and their design leads to
a better understanding of the application’s performance, and hence, leads to development
of better performance models. The Community Atmosphere Model (CAM) is the latest in
a series of global atmospheric models developed at the National Center for Atmospheric
Research (NCAR) as a community tool for NCAR and the university research community.
This work focuses on analyzing CAM and understanding the impact of different
architectures on this application. In the analysis of CAM, kernel coupling, which
quantifies the interaction between adjacent and chains of kernels in an application, is used.
All experiments are conducted on four parallel platforms: NERSC (National Energy
Research Scientific Computing Center) Seaborg, SDSC (San Diego Supercomputer
Center) DataStar P655, DataStar P690 and PSC (Pittsburgh Supercomputing Center)
Lemieux. Experimental results indicate that kernel coupling gave an insight into many of
the application characteristics. One important characteristic of CAM is that its
performance is heavily dependent on a parallel platform memory hierarchy; different
cache sizes and different cache policies had the major effect on CAM’s performance.
Also, coupling values showed that although CAM’s kernels share many data structures,
most of the coupling values are still destructive (i.e., interfering with each other so as to
adversely affect performance). The kernel coupling results helps developers in pointing
out the bottlenecks in memory usage in CAM. The results obtained from processor
partitioning are significant in helping CAM users in choosing the right platform to run
CAM. |
|---|