SeaSloth — Runoff mapping

Cost vs. ocean domain size

gen_rof_maps() builds the two mapping files CESM’s DROF component needs to move river discharge onto the ocean grid — a nearest-neighbor map and a smoothed one. Every case here regrids from the production GLOFAS mesh CIME registers as ROF_DOMAIN_MESH (21.6M elements), since that is the only source mesh a real case ever uses.

The ladder starts well below any real configuration — a 1° box, 144 ocean cells — and steps up roughly 4× per rung. The four box_* rungs are squares anchored at the Gulf of Mexico’s southwest corner, so the small end of the sweep is a zoom-in on the smallest real configuration rather than an unrelated patch of ocean. From the Gulf of Mexico up, every rung is a domain CROC actually builds.

What the curve says: Across the ladder, the destination grid grows 16,800× (144 → 2.4M ocean cells) while wall time grows 2.41× (77.73 s → 187.15 s). Cost is essentially flat up to the Caribbean — 805× more cells for 1.02× the time — and only then begins to rise, mildly and roughly linearly. What dominates the small end is fixed source-mesh work: every source-domain field (xc_a, yc_a, xv_a, mask_a, area_a, …) is ~2.2 GB and gets written into both output files, so a 1° box pays nearly the same toll as a basin. The destination grid only starts to matter at basin scale, and even there it is a factor, not a wall.

test_gen_rof_maps_domain — time vs. ocean domain

An ascending ladder of destination sizes at fixed 1/12° resolution: sub-real boxes first, then the domains CROC actually builds. Both nearest-neighbor and smoothed mapping files. Log-scale y, axis pinned to one decade so a flat curve reads as flat.

Production ROF mesh (21.6M elem)

Domain ladder — full numbers

Fixed 1/12° destination resolution, rmax = fold = 20 km.

Domain	Ocean cells	Mean	Peak RSS
1° box	144	77.73 s	12538.4 MB
2° box	576	73.75 s	12954.5 MB
4° box	2K	76.53 s	13368.9 MB
8° box	9K	74.18 s	13378.9 MB
Gulf of Mexico	34K	80.70 s	12588.4 MB
Caribbean	116K	78.90 s	12650.4 MB
N. Atlantic	576K	105.70 s	13235.8 MB
Indo-Pacific	2.4M	187.15 s	15801.6 MB

Cost vs. ocean resolution

The same call over one fixed footprint (Gulf of Mexico), so the source/destination overlap window and the set of contributing rivers are constant and only the destination cell count changes. This separates “how fine is the ocean grid” from “how much of the world does this domain touch” — and it confirms the domain ladder’s reading. Across the sweep, the destination grid grows 39× (4K → 146K ocean cells) while wall time grows 1.02× (79.69 s → 81.21 s). The ordering is not even monotonic — the spread is run-to-run noise.

test_gen_rof_maps_resolution — time vs. ocean resolution

Fixed Gulf of Mexico footprint, increasing destination resolution. Log-scale y, axis pinned to one decade.

Production ROF mesh (21.6M elem)

Resolution sweep — full numbers

Fixed Gulf of Mexico footprint, rmax = fold = 20 km.

Resolution	Ocean cells	Mean	Peak RSS
1/4°	4K	79.69 s	12563.0 MB
1/8°	15K	71.95 s	12589.0 MB
1/12°	34K	72.68 s	12591.9 MB
1/25°	146K	81.21 s	12606.0 MB