PCM Performance on the IBM SP
Winterhawk 1 and Winterhawk 2 Performance
Below are tables containing the performance of the Parallel Climate Model (version 1) on the IBM SP Winterhawk 1 and Winterhawk 2 nodes. Units are wallclock seconds per simulated day.
Winterhawk 1 Nodes
wallclock secs per simulated day
|
|
|
|
|
|
|
129.0 | 76.0 | 50.3 |
|
|
76.9 | 45.4 | 31.1 |
|
|
34.8 | 22.7 | 14.4 |
|
|
24.0 | 14.3 | 7.5 |
|
|
8.7 |
|
|
|
|
17.0 | 7.8 | 4.8 |
Winterhawk 2 Nodes
wallclock secs per simulated day
| Component | 16pes | 32pes | 64pes |
|
|
|
|
|
|
|
50.1 |
|
|
|
|
25.0 | 13.7 | 10.6 |
|
|
|
|
|
|
|
5.1 |
|
|
|
|
|
|
|
The performance characteristics of PCM on the IBM SP are well known. Namely:
Specific performance for
the PCM (version 2) coupler was produced by Tony Craig. Craig's
coupler PCM coupler performance summary .
Chris Ding and Yun He (NERSC) have
further analyzed the PCM (version 2) coupler performance, and have produced
improvements. Improvement summary .
The relative importance of I/O (on gseaborg at NERSC, the IBM SP WH1 machine) is shown in the following plot. During the integration, history and restart files are written at the end of every month for each component.

For comparison with other machines, see results presented in November 1999 at SC99 in Portland, Oregon: