Performance
Performance Roadmap — The Master List
Status: LIVING DOCUMENT. Every performance improvement found by the July 2026 measurement campaign (the five-way runtime comparison, the DB access-path matrix, the KV latency matrix, and three code audits), with its measured or estimated win and current state. Items move from “backlog” to “shipped” with receipts — numbers measured on the release artifact, per the benchmarks discipline. Last full revision: 2026-07-10, around the v0.4.1 release.
Shipped in v0.4.1
| Item | Measured win | Where |
|---|---|---|
TCP_NODELAY on every proxy, KV, and data-plane socket | ~200× per DB query (44 ms → sub-ms; every SELECT and connect paid a Nagle + delayed-ACK stall) | ephpm#161, litewire#3 |
| Hardware intrinsics restored in the PHP SDK (SHA-NI sha256, PCLMUL crc32, AVX2 base64 — disabled since the project began by one C++-only compiler flag) | sha256 2.7×, crc32c 2.2×; every HMAC, cache key, and ETag benefits | php-sdk#36 + build guard |
Quota-aware worker_count derivation + recycle default 500 → 10,000 | +24% worker throughput at container CPU quotas; recycle churn (one reboot per worker per ~0.25 s at 2k req/s) eliminated | ephpm#159 |
| litewire translate cache (LRU by query text) | 139× on repeated queries (38.6 µs → 277 ns per translate) | litewire#5 |
litewire prepare_cached + removal of the LIMIT-0 metadata probe | prepare round-trips halved | litewire#5 |
| litewire per-connection backends (WAL, real concurrent readers) | 52× on hot selects vs the reopen path; also fixes cross-connection transaction isolation | litewire#6 |
| Lazy-vhost negative-cache TTL 60 s → 2 s | regression fix: freshly deployed sites go live in seconds, not up to a minute | ephpm#164 |
WordPress worker per-request lifecycle (init/wp_loaded replay) | correctness unblock for worker-mode WooCommerce; ~2 ms/request cost | wordpress-worker v0.1.1 |
Release verification (post-fix DB matrix, sha256 ns/op on the shipped image, KV RESP-lane parity, before/after charts) accompanies the v0.4.1 release notes.
Backlog — v0.4.2 candidates, by estimated value
| Item | Est. win | State |
|---|---|---|
opcache.validate_timestamps = 0 in serve mode — deploys become events (ephpm deploy / ephpm cache reset), not per-request stat polling. Dev mode keeps instant reload; free win in immutable containers where files can’t change anyway | large on real apps (WordPress stats hundreds of files); biggest on container/network filesystems | design agreed: dev/serve split, startup log naming the contract, RESP-listener requirement surfaced, revalidate_freq=60 as the middle setting |
DB proxy buffering: BufReader per stream half + one coalesced write per response (today a 1-row SELECT costs ~24 syscalls; 4 are achievable) | ~25% of the ~190 µs proxy floor | audited with file:line (proxy audit era); not started |
Rust build tuning: lto = "fat", codegen-units = 1, panic = "abort", mimalloc as #[global_allocator] (currently: cargo defaults, system malloc) | 5–15% across all Rust hot paths | identified; not started |
| Lazy Envelope + interned header keys (worker dispatch Phase 1 — today five PHP arrays are built per request, most never read) | 30–50% of the 120 µs worker dispatch cost | draft exists: ephpm#160; needs containerized e2e + A/B before promotion |
| KV store internals (#142): coarse-clock LRU touch, O(1) eviction, TTL-only expiry index | ~15% of native GET (271 ns path) + latency-spike removal under memory pressure | measurement-justified by the KV matrix; not started |
| Query-stats overhead: cache metric labels per digest, short-circuit slow-path formatting | 5–10 µs per query on the SQLite lane | audited; not started |
| Proxy micro-trims: routing-loop double uppercase, fresh-connect challenge generation, pool checkout fast path | µs-class each | audited; not started |
| Worker dispatch Phases 2–3: per-worker SPSC slot ring, single-worker fast path, shared alloc sweep (#140) | toward the 60–80 µs/request target | worker dispatch fast path; gated on the Phase-0 flamegraph |
Gated on external milestones
| Item | Est. win | Gate |
|---|---|---|
| PGO for libphp | 10–15% across all PHP execution | upstream static-php-cli feat/pgo-v3 merging to v3 (php-sdk#35) |
| NTS prefork | removes the ZTS tax (~50% on allocation-heavy loops, 5–10% typical) | decision rule runs after PGO lands — nts-prefork |
| Turso engine (MVCC concurrent writes, native async I/O, sqld sidecar elimination) | removes SQLite’s single-writer wall; deletes an IPC layer from clustered writes | five gates, upstream GA first — turso-engine |
Unexplored — next audit sweep
Static-file serving internals (sendfile, pre-compressed asset cache), TLS session resumption configuration, PGO/BOLT for the Rust binary itself, KV compression threshold behavior, HTTP/2 tuning. No claims until measured.
Ground rules (how items earn their place)
- Measured before merged — every row above traces to a benchmark, a profile, or an audited code path with file:line. Estimates are labeled as estimates.
- Verified on the artifact — headline numbers are re-measured on the built release image before they appear in release notes.
- Guarded after shipping — silent-regression classes get CI guards (the SHA-NI symbol check, the opcache-enabled e2e) so wins can’t quietly evaporate.