The SWAT+ engine we run in production

Production engine pin

SWATGenX runs our fork at ship/v1.0-48.

Key takeaways

• SWATGenX runs our fork at ship/v1.0-48 in production — pinned per release, deployed 2026-08-01 (commit rafiei-vahid/swatplus@6520940 for auditors). Parallel execution is certified byte-identical to a single-threaded run of the same build in 18 of 18 runs — three trials at each of 2, 4 and 8 threads, in both parallel modes, over 935,130 compared values apiece — on a fixture that reproduces the defect 5/5 on the pre-fix binary.
• NetCDF output + the channel_sd print-filter cut output size 341× (695 MB → 2.0 MB) versus stock.
• The two engine fixes then attack compute: the measured engine core runs ~2.0× faster than stock, with byte-identical results — and the unified build is ship-gate certified output-equivalent to that core on plain models (≤ 1e-5).
• The unified engine adds opt-in OpenMP parallelism, PFAS transport, and a MODFLOW 6 coupling — all inert at the plain-model serial default, all deployed today.

What SWATGenX runs, and how far it is from stock SWAT+

"Production" should never be a vague word. SWATGenX runs our fork at ship/v1.0-48 — pinned per release — and it differs from the stock distribution in ways we can point to and measure. This page pins down exactly what we run and how far it is from the original.

It is the changelog and the integrator: a stock → fork-production ladder ending at the current unified engine, plus a table of every contribution and its state. The detailed proof for each lives on its own deep-dive page — here we show the cumulative effect, anchored to commits.

Every engine, pinned to a commit

Four engines are pinned for comparison. The stock baseline is swat-model/swatplus@5ccf6f0. Two fork rungs were each production in turn: 768f1d1 (NetCDF backend + channel_sd print-filter) and 247e95b (the same plus the two engine fixes). Current production is our fork at ship/v1.0-48 (commit rafiei-vahid/swatplus@6520940, deployed 2026-08-01): the 247e95b core consolidated with opt-in OpenMP parallelism (HRU land phase + routing wavefront), PFAS fate-and-transport, and a daily two-way MODFLOW 6 coupling. At its serial default all of that is inert, so a plain model runs exactly as before.

The three measured rungs were run on the same basin and window (Peace River HUC-8 (03100101) — 94,303 HRUs, 90 simulated days, ifx -O3 -ipo), each in its own native output mode — stock writes formatted-text channel_sd for all channels; the fork builds write gauge-filtered NetCDF — so the ladder reflects how each engine would actually be used, not an artificial common setting. The unified rung is deliberately not re-timed here: instead of a new benchmark, it carries a certification.

Before the 6520940 swap the candidate passed the full pre-production ship gate (scripts/ship_gate/gate.py): 75 checks PASS, 0 FAIL. Tier 1 on three plain benchmark models (M_09471300, MID_03100205, and MYAKKA_FERT_03100102 which runs scheduled fertilizer operations): backward-compatibility against the previous production engine byte-clean at exactly 0.00e+00, and parallel-vs-serial byte-identity at EXACT equality (0 differing cells over every output file) in BOTH parallel modes (HRU-parallel and full routing wavefront), at 2 and 4 threads, with 3 repeat trials per check — repeats included because a data race is probabilistic and a single clean run is not evidence. Tier 2 on the coupled SWAT+/MODFLOW-6 PFAS model: flow, nutrients and PFAS byte-clean against the golden reference, MODFLOW mass-balance discrepancy -0%, and coupled byte-identity for both parallel modes at 2 and 4 threads. Log: GenXAppData/ship_gate/runs/2026-08-01_v1.0-48-g6520940_vs_v1.0-45-g6423104_SHIP.log

• Baseline: swat-model/swatplus@5ccf6f0 (stock).
• Fork productions, in order: 768f1d1 (NetCDF + filter) → 247e95b (+ two engine fixes).
• Current production (deployed 2026-08-01): our fork at ship/v1.0-48 — unified engine, serial default, opt-in parallelism/PFAS/MF6.
• Each measured number on this page is sourced from the release manifest, not hand-edited.

Stock → fork productions → current unified engine

Peace River HUC-8 (03100101) — 94,303 HRUs · 90 simulated days.

Figure 1. Wall time and output size across the three measured rungs on Peace River HUC-8 (03100101) — 94,303 HRUs (90 simulated days). NetCDF + filtering crushed size; the two fixes crushed runtime. The current unified engine (6520940) is certified output-equivalent to the fastest measured rung on plain models, so it inherits these numbers at its serial default.

Table 1. Stock → fork productions → current unified engine, each pinned to a commit, with wall time, output size, and ratios versus stock for the measured rungs.

Unified engine (6520940): Not re-timed on this ladder. The ship gate certifies OUTPUT equivalence (worst-relative ≤ 1e-5) to the 247e95b rung on plain models at the serial default — that is a correctness result, not a wall-time one. The unified build is compiled with OpenMP, and an OpenMP build costs 1.20–1.24× at one thread against a serial-only build of the same core (measured twice on AWS c8a), so its serial wall time is not claimed to match the 247e95b rung. Opt-in multi-core scaling is measured separately on the parallel-engine page.

The contribution families move different axes. NetCDF + print-filtering (stock → 768f1d1) cut output ~341× (695 MB → 2.0 MB) while wall time changed only ~1.1× — because once output is small, the bottleneck is compute, not I/O. The two engine fixes (768f1d1 → 247e95b) then attack the compute: the same run drops to 227 s, ~2.0× faster than stock with byte-identical results. The unified engine (6520940) keeps that measured core — certified output-equivalent on plain models (≤ 1e-5) — and adds the opt-in capabilities: multi-core scaling (measured on the parallel-engine page), PFAS transport, and the MODFLOW 6 coupling.

Contributions and their state

Table 2. Every engine contribution: the axis it moves, its headline effect, its state (output/perf contributions have open upstream PRs; the parallel/PFAS/MF6 layers live on the open fork), and a link to its deep-dive evidence.

• Production is pinned per release — currently rafiei-vahid/swatplus@6520940 (deployed 2026-08-01), the unified engine: the measured 247e95b core plus opt-in OpenMP parallelism, PFAS fate-and-transport, and a MODFLOW 6 coupling, all inert at the plain-model default.
• Versus stock on a 94k-HRU basin, the measured engine core writes ~341× smaller output and runs ~2.0× faster, with byte-identical results; the unified build is ship-gate certified output-equivalent (≤ 1e-5) to that core on plain models.
• Output size and runtime are independent axes: NetCDF + filtering won on size, the two fixes won on speed, and both are in the shipped engine. The four output/perf contributions each have an open upstream PR against swat-model/swatplus (#213/#214/#219/#220); the parallelism, PFAS, and MODFLOW 6 layers live on the open fork's main branch and are documented on their own pages.