ICON scaling tests on JUWELS Booster at JSC

Recently a PRACE preparatory access application on the newly installed JUWELS Booster system at Jülich Supercomputing Centre (JSC) was accepted. The ICON team performed two strong scaling tests using two different horizontal resolutions: R2B7 (20km) and R2B9 (5km) to measure the performance of the code.

The tuning effort for optimization was spent for the high-resolution case R2B9 only and we reached a 75% efficiency using 512 GPUs compared to 256 GPUs.

Figure 1: Speed-up for two tests using different horizontal resolutions: R2B7(18km) and R2B9(5km). The dashed line represents a linear speed-up.
Figure 2: Wallclock time per time step at R2B7. The present data is shown in the blue curve. While previous tests results performed on the supercomputer Piz-Dant at CSCS in Switzerland were included for comparison. Note that on Booster, they are using the latest GPUs A100, while those used on Piz-Dant are from an earlier generation P100.

Through this exercise, we identified a few tuning parameters related to the code performance which will be very useful in the future work. Based on the present performance, we could simulate around 0.32 year per day using 600 GPUs at the resolution R2B9. In addition, simulations running  at even higher resolutions (e.g. about 1km) seem possible.

For more information on ICON visit https://code.mpimet.mpg.de/projects/iconpublic.


[post_grid id='651']