mirror of
https://github.com/opentibiabr/canary
synced 2026-08-16 06:26:09 -04:00
e99d283ebf
Improve Canary startup performance by reducing Lua/script loading cost, optimizing map parsing and cache construction, and lowering repeated lookup/allocation work during monster spawn startup. Measured impact: - Loaded modules and scripts wall time improved from about 54s to 1.516s best observed, with 2.997s in the latest captured run, about 94% to 97% less wall time. - Scripts::loadScripts dropped from 17.60% to 9.31% to 15.00% across later captures. - luaL_loadfile dropped from a dominant original hotspot to 1.60% to 2.80%. - getLuaBytecodeCacheEntry appeared at only about 0.49% to 0.51%, showing cache lookup is cheap compared to the old file-load path. - CanaryServer::loadMaps dropped from 65.59% after the Lua improvements to 57.68% in the latest capture. - IOMap::parseTileArea dropped from 59.58% to 53.87%. - MapCache::setBasicTile dropped from 24.96% to 20.50%. - phmap::try_emplace for cached basic tiles dropped from 7.25% to 0.99%. - MapSector::createFloor dropped from 7.17% to 5.76%. - SpawnMonster::startup dropped from 11.74% to 9.56%. - SpawnMonster::spawnMonster dropped from 10.64% to 9.13%. Main changes: - Add Windows RelWithDebInfo compiler and linker settings for useful release profiling symbols. - Add Lua startup telemetry options and bytecode cache configuration. - Make the Lua debug hook configurable so release profiling can avoid unnecessary hook overhead. - Improve Lua script loading by reducing filesystem metadata work, using bytecode cache entries, loading from memory when possible, and supporting packed cache data. - Add detailed startup timing logs for modules and datapack directories. - Add map-load reservations after reading the map header, including cache and container reservations based on map dimensions. - Add simple-tile and cache fast paths for common map-loading tile shapes. - Reduce shared pointer churn in the map cache path by using raw cache pointers internally where cache ownership is already guaranteed. - Reduce repeated sector and floor lookup work while parsing map tiles. - Improve zone indexing and copy behavior to avoid unnecessary set copies in spawn and tile insertion paths. - Optimize monster spawn startup by reusing resolved tile data and avoiding redundant lookups while preserving spawn behavior. - Move broad headers into pch.hpp where appropriate to keep benefiting from the precompiled header setup. - Add docs/maps/startup-loading-performance.md with profiling notes, decisions, results, and follow-up areas. - Add tests for map-download helpers and weighted monster spawn selection. - Add ignore rules for the Lua bytecode cache directory. Notes: - The profile captures are sampling profiles, so percentages represent sampled CPU share, not direct wall-clock speedup. - The latest 2.997s run appears to include local machine variance, but it remains far below the original 54s baseline. - Map loading is now the main remaining startup hotspot. - No behavior, map semantics, or spawn semantics are intended to change. |
||
|---|---|---|
| .. | ||
| startup-loading-performance.md | ||