mirror of
https://github.com/openfrontio/OpenFrontIO.git
synced 2026-07-05 09:50:48 +00:00
Evan
9e9c608053
## Summary Round 2 of GC-churn reduction, attacking the next tier of allocation sources found by the profiling harness from #4494. All changes are **behavior-preserving** — the simulation is bit-identical (final hash unchanged on three seeded runs). ### Changes | Site | Change | Churn target | |---|---|---| | `Player.units()` / `Game.units()` | Rest parameter → fixed-arity + array overloads (`units()`, `units(types[])`, `units(t1, t2?, t3?)`). The rest array was allocated on **every call** of one of the hottest functions in the sim. Spread call sites (`units(...Structures.types)`) now pass the array directly. `GameImpl.units()` builds one flat array instead of `Array.from().flatMap()` per-player intermediates. | ~18 GB | | `PlayerExecution` cluster flood fill | Results are plain `TileRef[]` in mark order instead of `Set<TileRef>` — the generation-stamped visited array already deduplicates, and consumers only iterate/measure. DFS stack reused across fills. | ~3.7 GB | | `SpatialQuery.bfsNearest` | Fused generation-stamped BFS with per-game scratch buffers (`WeakMap`-keyed, same pattern as `PlayerExecution`) instead of materializing a `Set` of the entire search area per query. Identical traversal and tie-breaking. | ~2.2 GB | | `NationWarshipBehavior` ship tracking | Single-pass loops instead of `filter().forEach()`; dropped defensive `Array.from(set)` copies (deleting the current entry while iterating a `Set` is well-defined). | ~1.4 GB | ### Results (Giant World Map, 400 bots, 12,000 ticks ≈ 20 game-min, seed `perf-default`) | Metric | Before | After | vs. pre-#4494 | |---|---|---|---| | Sampled allocations (incl. collected) | 59.2 GB | **37.8 GB (−36%)** | 97.7 GB (**−61%**) | | GC count / total pause | 1,076 / 1,830 ms | 772 / 1,442 ms | 1,682 / 3,313 ms | | Ticks/sec | 73 | **82** | 66 (+24%) | | Mean / p99 tick | 13.6 / 39.2 ms | 12.2 / 36.0 ms | 15.2 / 49.9 ms | `units()` no longer appears in the top-30 allocator list at all. The remaining leaders (possible round 3): the minimap pathfinding `Cell` pipeline (~8.5 GB), `diffPlayerUpdate`/`toFullUpdate` per-tick serialization (~4.6 GB), and iterator allocations (~3.3 GB). ## Determinism Final game-state hash unchanged on all three reference runs: - Giant World Map 12,000 ticks: `57830793797434300` ✓ - Giant World Map 2,000 ticks: `55125379638382860` ✓ - World 1,800 ticks: `32337437717390864` ✓ ## Test plan - [x] Full suite green (1,905 tests), including updated `units()` semantics tests (array overload, snapshot isolation, insertion order) - [x] Hash equality on 3 seeded headless runs (2 maps) - [x] Before/after 20-min GC benchmarks on the same commit base 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Fable 5 <noreply@anthropic.com>