ACCESS-NRI Model Scaling Repository: A collection of parallel scalability studies

ACCESS-NRI Releases
1

We are happy to announce the launch of the ACCESS-NRI Model Scaling Repository.

What is the ACCESS-NRI Model Scaling Repository?

The ACCESS-NRI Model Scaling Repository is a collection of Jupyter Notebooks that generate and display scaling data for ACCESS-NRI models.

We expect this data to be useful to both users and developers. It can be used to guide decisions about parallelisation layouts and CPU core counts when developing model configurations in order to ensure good balance between performance and parallel efficiency. It can also help identifying areas where models can be improved and optimised. Finally, the provided plots and tables can be used in NCMAS applications and other similar merit allocation schemes when requesting HPC resources.

Currently the repository includes scaling data for the following models:

We expect to regularly add new models to this list and update the existing notebooks when new versions of the models are available.

What should I do to be notified of any updates?

You can “watch” this topic (click the bell icon to the right - see full instructions here) to be notified by email of every update.

What is not included in these notebooks?

The notebooks do not include any type of resource usage estimate, like SUs per model year, as this information can strongly depend on the actual configuration being used. For ACCESS-NRI configurations, this information can sometimes be found in the release notes. If this information is not available, or you do not know how to obtain it, we suggest opening a new help request here on the ACCESS-Hive Forum.

Where do I ask questions?

Replies to this topic are disabled.

We welcome feedback and contributions here through the ACCESS-Hive Forum. Please create a help request in the Technical category and follow the guidelines for requesting help from ACCESS-NRI should you need it. Alternatively, if you are more comfortable with GitHub, you can open an issue in the scaling repository on GitHub instead of a help request on the Forum.

3

(deleted)

5

As mentioned in the above post, going forward we are not planning on releasing resource usage estimate along with the scaling studies.

We aware that this information is not always readily available, specially for configurations that are not officially supported by ACCESS-NRI or for which there are no stable releases yet. Considering that the NCMAS application deadline is tomorrow and following some requests from the community, we are providing bellow some resource usage estimate for several OM3 configurations.

Configuration Ocean Resolution Model Cost CPU Time Wall Time Years/Day OCN Procs Total Procs Queue Memory/Node
ACCESS-OM3 Global 25 km 1440 × 1152 × 75 17.2 kSU/yr 8394:49:29 03:11:06 7.8 2490 2704 Sapphire Rapids 57.5 GB
ACCESS-OM3 Global IAF 25 km 1440 × 1152 × 75 15.5 kSU/yr 7705:51:53 03:59:52 6.1 1685 1976 Sapphire Rapids 64.7 GB
ACCESS-OM3 Pan-Antarctic 4 km 4320 × 1442 × 75 98.4 kSU/yr 46548:35:16 13:51:00 1.76 2658 3432 Sapphire Rapids 77.3 GB
7

Thanks @micael. I guess the wall time is for 1 model yr the global models and 3 model months for panan, right?

8

@aekiss Thanks for spotting this. I’ve updated the table so that all times are for a full year.

9

For the upcoming NCMAS 2027 applications, the Software Transformation Team is planning to add scaling data for the following released configurations:

We will also be adding scaling information for the following configurations that are still under active development:

If you are planning to apply to NCMAS and the above scaling data does not fit your needs, please send me a private message or an email to see if we can help with that.

10

The new scaling data is now available for the following configurations:

The following configurations have also been updated or replace previous ones: