Notes from parallel BMI meeting IV

[mdpiper]

On 2024 Mar 19, we held a meeting (over Zoom) to discuss how BMI could work in a parallel environment. In attendance were @RolfHut, @hrajagers, @peckhams, @scrasmussen, @mcflugen, @rcabell, @PhilMiller, @hellkite500, and @mdpiper. (These are Mark's clumsy notes.)

Callback to Bert's branch describing a parallel BMI extension: https://github.com/csdms/bmi/tree/hrajagers/parallel-bmi

Soren:

• Using Bert's branch
• Serial coupling between WRF-Hydro and MODFLOW.
• Not yet parallel

Phil:

• Met with Nels. They are OK with calling parallel BMI functions collectively.
  • Bert met with Martijn on similar topic. It seems there are pros/cons to each technique.

Bert:

• We should discuss extensions. How should they work? Does core BMI object know about extension?

Rolf:

• Requests clarification on what is meant by parallelization.
• Will this functionality be an extension, not core? (Yes!)

Phil and Nels:

• Two primary use cases for NextGen.
• Using BMI C++ specification for all models.
• A few to a dozen models instatiated and coupled per cachement. Many cachements.
• Flow routing from cachements not yet BMI-ed.
• Currently one-way coupling. Tight coupling will be needed soon.
• Time direction

Rolf:

• grpc4bmi: ttps:https://grpc4bmi.readthedocs.io

Nels:

• Let's start simple, with something we can all agree on (e.g., collective call), but let's not lose sight of the possible need for something more complex for difficult problems.
  • Bert, Phil, Rolf: discussion of more complex approach

Bert:

• Are there issues with having core BMI and extension BMI in separate shared libraries? How will this work in parallel.