Unlocking the power of parallel computing: GPU technologies for ocean forecasting

Porter, Andrew R.; Heimbach, Patrick

doi:https://doi.org/10.5194/sp-5-opsr-23-2025

Articles | Volume 5-opsr

https://doi.org/10.5194/sp-5-opsr-23-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

Report:

Ocean prediction: present status and state of the art (OPSR)

https://doi.org/10.5194/sp-5-opsr-23-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 5-opsr

02 Jun 2025

| OPSR | Chapter 8.4

| 02 Jun 2025 | OPSR | Chapter 8.4

Unlocking the power of parallel computing: GPU technologies for ocean forecasting

Andrew R. Porter and Patrick Heimbach

Download

Final revised paper (published on 02 Jun 2025)
Preprint (discussion started on 26 Sep 2024)

Interactive discussion

Status: closed

RC1:
'Comment on sp-2024-32', Anonymous Referee #1, 01 Oct 2024

Good summary of the problems of porting particularly FORTRAN programs intended for CPUs and MPI to GPUs to gain advantages of performance and - ideally - reduced power consumption and heat generation.
If the FORTRAN codes are too big, too complex or too brittle to be changed, then the interface to the distributed/parallel processing capability requires considerable work to be sufficiently general to handle a range of application codes on one hand, and a range of distributed/parallel processing systems on the other.
This difficulty - and possible solutions - should be discussed in the paper and would increase greatly the value of the contribution. These solutions range from using software other than MPI, defining a new interface (with middleware) between FORTRAN codes and parallel/distributed processing at both node and processor levels or revising the FORTRAN codes. Developments in processor technology may introduce other opportunities beyond the current generation of GPUs.
Also, it would appear that the problem is being addressed by OpenMPI and that they have a version of MPI that takes advantage of GPUs.

Citation: https://doi.org/10.5194/sp-2024-32-RC1
- AC2: 'Reply on RC1', Andrew Porter, 16 Dec 2024
  
  Many thanks for the review. Below we address the reviewer’s comments and suggestions (reviewer’s comments in italics, our responses follow):
  "This difficulty - and possible solutions - should be discussed in the paper and would increase greatly the value of the contribution. These solutions range from using software other than MPI, defining a new interface (with middleware) between FORTRAN codes and parallel/distributed processing at both node and processor levels or revising the FORTRAN codes. Developments in processor technology may introduce other opportunities beyond the current generation of GPUs."
  We concur with the reviewer regarding challenges and solutions, however we think that the manuscript does describe various solutions, i.e., we address the various ways in which OOFS can make use of GPUs, ranging from the addition of OpenACC/OpenMP directives to existing Fortran through to complete re-writes in C++, Python and Julia. We are unsure in what way we should extend the discussion.
  "Also, it would appear that the problem is being addressed by OpenMPI and that they have a version of MPI that takes advantage of GPUs."
  As you say, GPU-aware MPI implementations are important and we have extended the text to make this clear.
  
  Citation: https://doi.org/10.5194/sp-2024-32-AC2
RC2:
'Comment on sp-2024-32', Mark R. Petersen, 30 Nov 2024

The comment was uploaded in the form of a supplement: https://sp.copernicus.org/preprints/sp-2024-32/sp-2024-32-RC2-supplement.pdf

Citation: https://doi.org/10.5194/sp-2024-32-RC2
- AC1: 'Reply on RC2', Andrew Porter, 16 Dec 2024
  
  Many thanks for the useful review and the additional technical information on the ocean component of E3SM that is being developed. It will be very interesting to see how your work there progresses. We have added this information to the text.
  The reviewer comments:
  "The AI/ML revolution is causing much greater uncertainty, as it may displace the numerical methods we’ve used for 70 years. I think this point deserves a few more sentences in the text."
  This is a good suggestion. We've added some text about this issue.
  To respond to the remaining points:
  L43: We have updated the Top500 reference (and associated text) to use the latest list as of Nov 2024;
  L59: Thanks for pointing out that we'd missed describing inter-processor communication and the additional complexities this brings for GPUs. We've added a paragraph explaining the need for this communication and the implications for GPU.
  L98-112: We've added a little more detail here.
  The Strauss et al. reference has been updated (thanks).
  
  Citation: https://doi.org/10.5194/sp-2024-32-AC1

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

ED: Publish subject to minor revisions (review by editor) (18 Jan 2025) by Jay Pearlman

AR by Andrew Porter on behalf of the Authors (04 Feb 2025) Author's response Author's tracked changes Manuscript

ED: Publish as is (10 Feb 2025) by Jay Pearlman

AR by Andrew Porter on behalf of the Authors (17 Feb 2025)

Short summary

Numerical ocean forecasting is a key part of accurate models of the Earth system. However, they require powerful computing resources, and the architectures of the necessary computers are evolving rapidly. Unfortunately, this is a disruptive change – an ocean model must be modified to enable it to make use of this new computing hardware. This paper reviews what has been done in this area and identifies solutions to enable operational ocean forecasts to make use of the new computing hardware.