swatplus-engine-acceleration

Reproducibility package for:

Accelerating a Regional Hyper-Resolution SWAT+ Model Without Changing the Science: a Reproducible Profiling-to-Parallelization Methodology and a Routing-Aware Shared-Memory Wavefront V. Rafiei (submitted to Computers & Geosciences).

This repository contains everything needed to reproduce the benchmarks in the paper: the benchmark model, the build recipe for the optimized/parallel SWAT+ engine, the timing harness, the measured results, and the figure-generation script.

You do not need AWS. The paper's timings were collected on dedicated AWS instances only to get quiet, uncontended hardware, but the benchmark itself is hardware-agnostic — run bench_scaling.sh on any machine and you get that machine's scaling curve. The provisioning scripts we used are in aws/ purely as a convenience; they are not part of the reproduction path.

What's here

Path	What it is
BUILD.md	Exact engine fork + per-optimization commits + the ifx build recipe
peace-river-03100101-model.7z	The benchmark model (Peace River, HUC-8 03100101; 57,998 HRUs / 8,181 channels). Unpack with 7z x.
bench_scaling.sh	Cross-hardware scaling sweep: stock + final engine, N = 1…24 threads, best of two
bench_full.sh	Full study: serial ladder + text/NetCDF I/O + thread-count byte-identity sweep
make_figures.py	Regenerates the paper figures into figures/
make_tables.py	Recomputes the scaling tables (Table 2 / S2 / S3) directly from results/*.csv
results/*.csv	The measured wall times behind the paper's tables and figures
aws/	Optional. The scripts we used to provision dedicated cloud instances

The engine (built from commits, not shipped as a binary)

Every optimization is a single commit in the engine fork, so the cumulative ladder is auditable commit by commit. See BUILD.md. The benchmark binaries are not committed here — they are reproducible from those commits with ifx, and binaries are compiler/CPU-specific.

Quick start

# 1. unpack the model
7z x peace-river-03100101-model.7z

# 2. build the engine binaries (see BUILD.md) and place them in ./bins/:
#    r0_<commit>  (stock)   r6_<commit>  (final serial)   r6clean (final, OpenMP)

# 3. run the scaling sweep on YOUR machine
bash bench_scaling.sh            # writes results.csv

# 4. regenerate the figures
python3 make_figures.py          # writes figures/Figure_*.png

# regenerate the scaling tables straight from the measured CSVs (no embedded numbers)
python3 make_tables.py            # reproduces Table 2 / S2 / S3

A note on the numbers

Absolute wall times and speedups are hardware-specific — they depend on CPU microarchitecture, core count, memory bandwidth, and compiler version. The paper reports three dedicated-instance microarchitectures (AMD EPYC "Turin", Intel Xeon "Granite Rapids", AMD EPYC "Rome"); your machine will differ. The shape of the result — a wide, abundantly parallel land phase followed by a memory-bound, serially-bounded routing tail — is the portable finding.

License

The benchmark harness and scripts here are released under the MIT License. The SWAT+ engine itself is governed by the upstream swat-model/swatplus license.