Oceaningans novel advection scheme

@navidcy shared this and I thought it might be of interest to the community

This novel advection scheme sounds fascinating. Anyone care to share details or make a comment? Will it be adopted in other ocean models?


Here’s the high-resolution image included in the blog post.

Surface vorticity in the Kuroshio region simulated with Oceananigans @ 1/12th and 1/48th resolution.

(Download, zoom in and explore submesoscale instabilities and what not…)

1/12th near-global simulation using 4 GPUs Nvidia V100s achieving 1 simulated year per day (on NCI that costs ~3kSU per simulated year)

1/48th near-global simulation using 32 GPUs Nvidia A100s achieving 15 simulated days per day (on NCI that costs ~100kSU per simulated year)

[Disclaimer: These are not perfect! There is a lot of work to be done… Forcing files, interpolations, bathymetry tweaking, etc. But still the sheer computation and the eddies look cool so I shared. :wink: If you want to work with us then reach out.]


Hm… seems like discourse chops down the file to a much smaller size. Download the full 39MB image from dropbox.

The advection is done via a WENO scheme (with some extra ingredients)! We are currently writing this up.

The Kuroshio seems to meander less and persist for longer in the 1/48th simulation. Is that borne out in the analysis?

The straight-up performance numbers are very very impressive.

Also 39MB file is pretty large. I’ve halved the size in both dimensions and saved as JPG, we’ll see it if lets that through without alteration

We haven’t yet done any/much analysis of this simulation yet…

See the disclaimer I put up when I first post the image.

But it’s a very good point you are raising @Aidan.

1 Like

@navidcy it would be great if you submitted this to JAMES.