René's URL Explorer Experiment

Title: http: avoid dictionary-mode objects in responses by mcollina · Pull Request #66420 · nodejs/node · GitHub

Open Graph Title: http: avoid dictionary-mode objects in responses by mcollina · Pull Request #66420 · nodejs/node

X Title: http: avoid dictionary-mode objects in responses by mcollina · Pull Request #66420 · nodejs/node

Description: setHeader() stored headers in a { __proto__: null } literal and OutgoingMessage let EventEmitter create the same literal for its listener table. V8 creates null-prototype literals in dictionary mode, so every response paid for a hash table allocation on the first setHeader()/on() call, for runtime-miss stores (the profile shows StoreIC::Store, Factory::NewStoreHandler and TieringManager::NotifyICChanged in steady state) and for dictionary lookups on every header access, including the for...in in _storeHeader(). This PR uses a fast-mode object with an empty null-prototype chain for the headers (same keys, same enumeration, no inherited properties), and presets the listener table with the common events, as Readable/Writable already do. Hello-world server (res.setHeader('Content-Type', ...) + res.end('Hello World')), server pinned to one core, wrk -t2 -c50 on two other physical cores, 4 interleaved rounds: req/s user µs/req sys µs/req alloc B/req main 40.9k 15.0 9.4 4374 this PR 50.0k 11.3 8.7 3905 The _events preset accounts for ~2% of that; the header container for the rest. With 8 request headers and a handler that reads two of them: 36.2k → 43.2k req/s. Responses are byte-identical. How much the dictionary-mode container costs depends on the response pattern (same harness, 2 rounds each): handler main this PR setHeader('Content-Type') + end(body) 41.7–42.4k 51.3–51.8k setHeader('Content-Type') + setHeader('Content-Length') + end(body) 47.1–49.6k 50.0–51.7k writeHead(200, { ... }) + end(body) (no setHeader(), benchmark/fixtures "normal") 46.1–50.0k 50.2–50.9k The existing benchmark/http fixtures either set Content-Length through setHeader() or pass every header to writeHead(), so they only show the smaller gains. The second commit adds a setHeaderImplicit case to http/set-header.js (res.setHeader('Content-Type', ...) + res.end(body), framing left to end()). benchmark/compare.js --runs 10 --filter set-header http: confidence improvement accuracy (*) (**) (***) http/set-header.js duration=5 res='normal' benchmarker='wrk' 2.53 % ±4.03% ±5.61% ±7.86% http/set-header.js duration=5 res='setHeader' benchmarker='wrk' ** 6.13 % ±4.03% ±5.55% ±7.61% http/set-header.js duration=5 res='setHeaderImplicit' benchmarker='wrk' *** 16.46 % ±5.10% ±7.06% ±9.78% http/set-header.js duration=5 res='setHeaderWH' benchmarker='wrk' 4.78 % ±5.76% ±7.89% ±10.75% simple.js, headers.js and incoming_headers.js are unchanged within noise. —- Ai generated, humanly reviewed.

Open Graph Description: setHeader() stored headers in a { __proto__: null } literal and OutgoingMessage let EventEmitter create the same literal for its listener table. V8 creates null-prototype literals in dictionary mod...

X Description: setHeader() stored headers in a { __proto__: null } literal and OutgoingMessage let EventEmitter create the same literal for its listener table. V8 creates null-prototype literals in dictionary mod...

Opengraph URL: https://github.com/nodejs/node/pull/66420

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width