Poor URLPattern performance

[kitsonk]

I was looking at the performance of my oak router versus my acorn router. Oak uses pathToRegex while acorn uses URLPattern. The implementation of URLPattern in Deno is super non-performant compared to pathToRegex and is making any routers that leverage URLPattern quite noticeably non-performant against other solutions.

By my calculations, URLPattern is ~150x slower than pathToRegex when testing/executing.

To reproduce:

benchmarks.test.ts

import { pathToRegexp } from "https://deno.land/x/[email protected]/index.ts";

const urlPattern = new URLPattern("/", "http://localhost/");

Deno.bench({
  name: "URLPattern",
  fn() {
    if (urlPattern.test("http://localhost/")) {
      true;
    }
  },
});

const regexp = pathToRegexp("/");

Deno.bench({
  name: "pathToRegexp",
  fn() {
    if (regexp.test("/")) {
      true;
    }
  },
});

And then:

❯ deno bench benchmarks.test.ts
cpu: Apple M1 Pro
runtime: deno 1.35.1 (aarch64-apple-darwin)

file:///Users/kitsonk/github/acorn/benchmarks.test.ts
benchmark         time (avg)             (min … max)       p75       p99      p995
---------------------------------------------------- -----------------------------
URLPattern         2.33 µs/iter     (2.23 µs … 2.93 µs)   2.38 µs   2.93 µs   2.93 µs
pathToRegexp      14.23 ns/iter   (14.01 ns … 26.11 ns)   14.3 ns  16.23 ns  17.06 ns

[bartlomieju]

Thanks for flagging it Kit, we actually discussed it last week. We'll look into optimizing it.

[bartlomieju]

Quick profile from V8 shows that op_urlpattern_process_match_input is one of the culprits - it is a "slow op" and additionally produces a lot of garbage that V8 has to clean up.

URLPattern.exec will suffer from the same problem.

[kitsonk]

Yeah, I noticed both .exec and .test have a similar performance profile delta between the two approaches.

[littledivy]

[marvinhagemeister]

Was looking at a couple of server boot Fresh traces and URLPattern shows up there pretty prominently. Whilst the matching itself is slow as demonstrated by the original comment, constructing URLPattern instances is equally slower than just winging it with a regex.

[bartlomieju]

@littledivy is actively looking into improving performance of URLPattern.

[marvinhagemeister]

For completeness: The pattern to regex parser I used can be found here https://github.com/denoland/fresh/compare/url-matcher#diff-7d5020e43e696993c4dee3edff778d85149f8165379a8a480824b53dfff7bbdeR1133-R1210 . Mostly used it as a starting point to see how much of a difference there is between a regex and URLPattern so it's probably not complete or spec complient.

[lucab]

On top of raw performance timing, after profiling some isolates in production we suspect that this is also causing a lot of memory pressure. One possible culprit is the StringOrInit input argument:

• it is a quite large enum (with unbalanced variant sizes) and has to go through untagged serialization/deserialization.
• in most cases it is used to hold a simple URL String.
• it normally appears in the hot-path of requests handling (possibly even in hot-loops).

[kitsonk]

Still an issue. While URLPattern did go from 166x slower to only 23x slower, it is still a challenge (as well has the highly variable p995). This really impacts any framework that attempts to do routing via URLPattern:

❯ deno task bench
Task bench deno bench
cpu: Apple M1 Pro
runtime: deno 1.40.1 (aarch64-apple-darwin)

file:///Users/kitsonk/github/acorn/routing.bench.ts
benchmark         time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------ -----------------------------
URLPattern       334.65 ns/iter   2,988,229.5   (287.25 ns … 5.19 µs) 307.73 ns 704.6 ns 5.19 µs
pathToRegexp      14.63 ns/iter  68,362,672.9   (14.04 ns … 56.73 ns) 14.82 ns 19.48 ns 21.65 ns