SeaSloth MOM6 Scaling Report

Location-dependent scaling sweep on Derecho. · Raw data (JSON)

Throughput by Geographic Location

Compset: 1850_DATM%NYF_SLND_SICE_MOM6%REGIONAL_SROF_SGLC_SWAV — machine: derecho, queue: develop — date range: 2020-01-01, 2020-01-31

Each domain is 100×100 grid points, flat 1000 m bathymetry, 10 vertical levels, run at NTASKS_OCN=20 and 40 on Derecho's develop queue (~24-day simulation). Does MOM6/CESM throughput vary by where on Earth the domain sits?

Note: DT (baroclinic timestep) is auto-selected by CrocoDash based on grid spacing — a coarser grid allows a larger CFL-stable DT, which reduces wallclock time independently of geographic location. All domains here use a matched DT=600 s so the comparison reflects location effects rather than timestep artifacts. Increasing or decreasing DT has a direct proportional effect on throughput.

Throughput (sim-years/day) by domain and NTASKS_OCN

Caribbean N=20: 76.0 y/dayCaribbean N=40: 115.3 y/dayEquatorial Pacific N=20: 71.0 y/dayEquatorial Pacific N=40: 115.0 y/dayArctic N=20: 64.0 y/dayArctic N=40: 93.2 y/dayAntarctic N=20: 69.0 y/dayAntarctic N=40: 115.8 y/day N=20N=40
CaribbeanEquatorial PacificArcticAntarctic

Cross-domain comparison — throughput (sim-years/day)

DomainN=20N=40
Caribbean76.02115.30
Equatorial Pacific71.04115.01
Arctic64.0193.18
Antarctic69.04115.78

Caribbean

100x100 (0.1deg, ~11km), centered ~20N 75W, flat 1000m, nk=10

NTASKS_OCNThroughput (sim-years/day)Total Run TimeOCN Cost (pe-hrs/sim-year)Parallel efficiency (vs N=20)
2076.021.2 min40.4100%
40115.3047.2 s26.676%

Equatorial Pacific

100x100 (0.1deg, ~11km), centered ~0N 140W, flat 1000m, nk=10

NTASKS_OCNThroughput (sim-years/day)Total Run TimeOCN Cost (pe-hrs/sim-year)Parallel efficiency (vs N=20)
2071.041.3 min43.2100%
40115.0147.3 s26.781%

Arctic

100x100 (10km, polar-stereo EPSG:3995), 1000x1000km cap on N pole, flat 1000m, nk=10

NTASKS_OCNThroughput (sim-years/day)Total Run TimeOCN Cost (pe-hrs/sim-year)Parallel efficiency (vs N=20)
2064.011.4 min48.0100%
4093.1858.4 s33.073%

Antarctic

100x100 (36km, polar-stereo EPSG:3031), 3600x3600km cap on S pole (boundaries ~67-74S, widened to clear GLORYS's -80S coverage limit), flat 1000m, nk=10, DT=600s matched to other domains

NTASKS_OCNThroughput (sim-years/day)Total Run TimeOCN Cost (pe-hrs/sim-year)Parallel efficiency (vs N=20)
2069.041.3 min44.5100%
40115.7847.0 s26.584%

Reference: Bahamas Full NTASKS Sweep

100x100 (0.1deg), centered ~30N 75W, flat 1000m, nk=10 — queue: develop (N≤40); main (N=80)

Throughput vs. NTASKS_OCN (1–80 cores)

x-axis: NTASKS_OCN

1: 4.9 y/day4.9 y/day2: 13.0 y/day13.0 y/day4: 15.4 y/day15.4 y/day5: 17.9 y/day17.9 y/day10: 34.2 y/day34.2 y/day20: 69.3 y/day69.3 y/day40: 120.5 y/day120.5 y/day80: 148.5 y/day148.5 y/day 124510204080
NTASKS_OCNThroughput (sim-years/day)Total Run TimeOCN Cost (pe-hrs/sim-year)Parallel efficiency (vs N=1)
14.9024.2 min627.2100%
213.049.1 min235.6133%
415.457.7 min198.879%
517.916.6 min171.673%
1034.223.5 min89.870%
2069.271.7 min44.471%
40120.4858.9 s25.561%
80148.4647.8 s20.738%