Skip to content

PERFORMANCE.md

Latest commit

 

History

History
67 lines (53 loc) · 3.04 KB

PERFORMANCE.md

File metadata and controls

67 lines (53 loc) · 3.04 KB
Raw

Performance Notes

Build flags (take the free speed)

go build -pgo=auto -trimpath -ldflags "-s -w" ./cmd/asyncao   # release
GOAMD64=v3 go build ...                                        # AVX2 builds (2013+ CPUs)

-pgo=auto picks up default.pgo at the repo root. Ship both a v3 and a baseline build; v3 refuses to start on pre-Haswell CPUs.

Memory budget

cmd/asyncao sets debug.SetMemoryLimit(256 MiB) at startup. The byte budgets stack under it: T1 64 MiB textures + T2 128 MiB bytes + working set. GOMAXPROCS stays default — the netpoller already covers blocking I/O; the old "+2 for I/O" advice is a myth.

PGO capture

default.pgo is a CPU profile from a scripted courtroom session. To re-capture after significant changes:

# 1. run the client with pprof enabled
./asyncao -debug &        # pprof on localhost:6060
# 2. drive a session: connect, char select, ~5 minutes of paired IC messages
#    (scripts/pgo-session.md documents the manual script)
# 3. capture while the session is active
curl -o default.pgo "http://localhost:6060/debug/pprof/profile?seconds=120"
# 4. rebuild with -pgo=auto and commit the new profile

A profile captured during real courtroom traffic (typewriter + animations + fetch/decode) optimizes the paths that matter: decode loops, render copies, LRU lookups.

Profiling a live client

./asyncao -debug
go tool pprof http://localhost:6060/debug/pprof/profile?seconds=30   # CPU
go tool pprof http://localhost:6060/debug/pprof/heap                 # heap
go tool pprof http://localhost:6060/debug/pprof/allocs               # alloc churn

Watch for:

  • any allocation inside render.Viewport.Render / Update (the gate is 0),
  • runtime.cgocall overhead spikes (batching texture ops beats chatty calls),
  • GC pause p99 in the 1 Hz sampler (budget < 2 ms; the 256 MiB soft limit plus low allocation rates keep cycles rare and short).

Hot-path inventory

Path Technique
Format resolution atomic snapshot, pooled candidates, 1 alloc — and the unlearned path reads a generation-cached format table (no prefs lock, no rebuild)
Cache hits lock-free-ish LRU (internal lock, no wrapper), atomics for stats
Fetch dedup singleflight; 404 LRU; per-host backoff off the hot path
Decode worker pool, size-classed pixel pools, SIMD libwebp via CGO
Texture upload budgeted per frame; live message bypasses the budget
Typewriter rasterize once per message; reveal = src-rect width per frame
Animations precomputed delay tables; frame advance is an index bump
Render zero allocations steady-state (cgo-escape pitfalls: never take the address of a stack rect for Renderer.Copy — use a reused field); texture pages are generation-cached per layer, so steady frames do zero LRU lookups
Animated decode composition canvas + DisposalPrevious snapshot from the pixel pool — allocations limited to the output frames
HTTP transport 3 s response-header timeout fails stalled hosts before the 5 s deadline, freeing the per-host connection slot