I finally successfully built JEDI-MPAS without Containers after falling back to Eigen 3.3.9. Appreciate this forum’s help regarding my Eigen issue. Now I’m ready to start running JEDI-MPAS. For background, I am running on a MacBook, using VirtualBox with 6 cores, and will never be running in a (huge #)-core HPC system environment. I simply want to be able to assimilate a local sounding and some surface reports into MPAS running in regional mode (for example the size of Texas (US state - never doing global simulations). I looked at DART for the data assimilation component briefly, but realized that it is strictly for fairly large ensembles, again something I won’t be doing. This led me to JEDI for the data assimilation.
So my question is given all this, is it mandatory to install/run Singularity to run JEDI-MPAS (the tutorials or general execution)? An associated question: do I have to use a Container (again for the tutorials or for general execution)? JEDI-MPAS compiled just fine for me without Containers.
I’m trying to stay as lightweight as possible in all this: quite small MPAS simulation areas/no ensembles eliminating the need for 100’s of cores and very small data segments for assimilation. Let me know please. TIA.
Neither Singularity nor any other container is necessary to run jedi-mpas.
A caveat on your lightweight strategy: The most mature assimilation technique for jedi-mpas is EnVar. It is also, at this point, the only one covered in the tutorials and documentation. EnVar requires an ensemble of forecasts for use in the background covariances.
That’s really good news regarding Singularity/containers In the tutorials, it seems they all involve containers and Singularity. Would there happen to be any tutorials that I could follow that don’t involve Singularity/containers so I can spin up on how to run things? Point taken on EnVar - however, what assimilation technique, within jedi-mpas, would you recommend, for my purposes, i.e., assimilating a set of surface obs and a sounding or two? TIA.
The reason for Singularity/containers in the tutorials is in order to establish a working environment. In short, you will need to establish the set of libraries on your laptop. It is the same set of libraries for all bundles (including fv3). Unfortunately that is not my area of expertise. @maryamao might be able to give more information.
JEDI-MPAS developers are actively working on getting 3dvar (without ensembles) to work effectively. There is a lot of work that goes into properly training a climatological B. @byoung-joo might be able to say more, but you may have to choose between training your own climatological B for your domain in order to use 3dvar or running ensembles of short-range forecasts (again, on your domain) in order to use 3denvar.
To my knowledge, we do not yet have a tutorial on how to train the climatological B because this is an active area of development.
Although most of our work with JEDI-MPAS so far uses global meshes, we have used regional meshes in alpha testing with success. So that in itself should not pose problems. If you intend to cycle with alternating forecasts and DA, then you will need to decide how to update your lateral boundary conditions, which is not currently enabled within JEDI-MPAS.