The grid class requires a large amount of memory and time to set up

[b-barton]

The grid class needs re-working to reduce memory requirements. At the moment it works and is fine for smaller parent and child grids but it doesn't scale well to large grids like UK500 and NAARC (which uses a masked global domain). This means these larger grids need unreasonable amounts of memory allocation on JASMIN (384Gb).

The solution is to start introducing xarray for reading in grid information. PyBDY only needs parent and child grid information along the boundaries. This is:
Parent: bdy +rim +9 surrounding points +2 for calculating grid variables e1, e2, glamu, levels with partial setp etc.
Child: bdy +rim +2 for calculating grid variables.
At the moment is gets grid information for the whole domain for both. With xarray we can work out where we will need grid information and then only extract it along required boundaries (this is similar to what pybdy does for src data with nemo_bdy_extr_assist.get_ind()).

[b-barton]

Linked to #170

[b-barton]

1. load mask/bathy
2. load hgr for dst and src
3. get bdy chunk points
4. find max +1 and min -1 for dst loading and src loading
5. subset hgr variables for each chunk
6. load zgr data for each chunk

[b-barton]

I've removed the deepcopy of grid from Extract and it makes 20% reduction to the max memory requirement for an example grid.

[b-barton]

I've found indices for the bdy points and subset the hgr and zgr for dst and sc. The bdy points are selected in chunks so hgr and zgr now have chunks. This new way of interacting with hgr and zgr needs integrating into the rest of the code.

1. z at bdy points - done
2. tide extract - done
3. Extract - done

[b-barton]

The changes work and the code runs. Chunking the grid data gives a further 40% reduction in memory requirement.
The regression test results have changed a bit which is unexpected so I'll work out why that is happening. Some of the depths have changed in sco test so probably related to that.

As a memory performance guide the dst domain is 296 x 489 x 51 points and the source used here is 241 x 271 x 75 points (y, x, z).

[b-barton]

I've fixed the bugs and now the regression tests pass and the output from my example looks the same as for the master. Total memory saved is 54% for this example.

[b-barton]

I've got the NAARC domain running with pyBDY. This is the memory profile from master (before memory improvements).

[b-barton]

There is a 10x improvement in memory requirement for the NAARC domain with the memory improvements. There is also a 6x improvement in run time (time in seconds on graph).

As a memory performance guide the NAARC domain is 3605 x 4320 x 75 points and the source used here is 1206 x 1440 x 75 points (y, x, z).

[b-barton]

There is an error that comes up for some domains when there is a small (a few points) boundary chunk that no (parent?) points are found in. As a minimum, this should be replaced with a more useful error message.

File "/home/users/benbar/pyBDY/src/pybdy/nemo_bdy_gen_c.py", line 104, in init
msk = np.pad(bdy_msk, ((1, 1), (1, 1)), "constant", constant_values=(-1))
ValueError: operands could not be broadcast together with remapped shapes [original->remapped]: (2,2) and requested shape (0,2)