Replies: 3 comments 6 replies
-
Hi @jklymak! I'll let @navidcy and @glwagner comment on the powerful algorithm since they developed the example in #1125.
Was just curious about your MPI concerns: how large are the simulations you'd like to run? Oceananigans can do 100~150 million grid points on a 32 GB GPU (e.g. a V100) which covers some simulations that tend to require multiple CPUs. Of course, if your simulations are larger or if large GPUs aren't super available then MPI + CPUs is the way to go. @tomchor has also raised a similar concern about MPI (#1234 (comment)). MPI support is technically under-development (PR #590) but the priority keeps getting pushed... I see you also came across @whitleyv's immersed boundary work (#1036). Hopefully in some months Oceananigans will have MPI and immersed boundary support! Sorry to derail this thread! Happy to continue discussing in a new thread to keep this one on-topic. |
Beta Was this translation helpful? Give feedback.
-
That’s a good point! Thanks. |
Beta Was this translation helpful? Give feedback.
-
@jklymak raises some good points. In terms of the power method, I agree tha tit can be made clearer. I think the basic idea is that we want to get the growth rate and spatial structure in the context of the linear theory. I think this should be equivalnet to just solving the linearized eqations. I wonder if making this connection might help? I am also keen to help with the MPI issue. The plan is to start testing with ShallowWaterModel since there is no elliptic problem to solve and hence a bit easier. This is for when people have time of course. |
Beta Was this translation helpful? Give feedback.
-
I was looking over the documentation for oceananigans. It is looking very exciting, though I don't imagine I can seriously use it until MPI and immersed boundaries work.
I found the discussion of the "Powerful algorithm" in https://clima.github.io/OceananigansDocumentation/stable/generated/kelvin_helmholtz_instability/ a bit more obscure than perhaps necessary. It seems that you are successively filtering the solutions so that you emphasize the fastest growing mode, but keep the amplitude of the velocities in a regime that will let the model stay linear. What I was missing is a clearer motivation of "why". Can you include a couple of sentences about what the end goal of this would be, and why it is not achievable just running a free simulation? I think its something like:
but I really had to think to dredge that up, and I'm not sure if I am even correct.
Beta Was this translation helpful? Give feedback.
All reactions