Use llhttp wasm build #575

dnlup · 2021-02-27T13:39:28Z

In this PR I'll try to:

setup a branch in https://github.com/dnlup/llhttp/tree/undici_wasm which defines some npm build scripts to make a wasm build
setup a script to run the build in this repo and save the build output
implement a minimal parser that is compatible to use it as a base class in undici parser (I am still undecided if it should live in llhttp or here)
test that the parser is a drop-in replacement (hopefully) and performance degradation is minimal

~~This is still messy, I'll get there eventually.~~

Final tasks:

implement lenient option
~~implement headersTimeout option~~
implement maxHeadersSize option

Fixes #22

ronag

You only need to implement RESPONSE for client.

lib/llhttp/parser.js

mcollina · 2021-02-28T10:07:08Z

The goal for this activity should be to verify that we suffer minimal slowdowns from using wasm instead of native.

lib/llhttp/parser.js

ronag · 2021-02-28T10:25:40Z

We also probably want to reduce the number of calls across the js <-> wasm border so we might want to modify the c level code as well. Let's just get it working and passing tests first and then we optimize.

mcollina · 2021-02-28T10:44:36Z

I agree!

lib/llhttp/parser.js

codecov-io · 2021-02-28T15:20:30Z

Codecov Report

Merging #575 (66af633) into master (3d002c1) will decrease coverage by 1.36%.
The diff coverage is 92.71%.

@@            Coverage Diff             @@
##           master     #575      +/-   ##
==========================================
- Coverage   99.36%   97.99%   -1.37%     
==========================================
  Files          16       17       +1     
  Lines        1414     1547     +133     
==========================================
+ Hits         1405     1516     +111     
- Misses          9       31      +22

Impacted Files	Coverage Δ
lib/llhttp/parser.js	`92.36% <92.36%> (ø)`
lib/core/client.js	`98.76% <100.00%> (-0.38%)`	⬇️
lib/llhttp/constants.js	`100.00% <100.00%> (ø)`
lib/core/util.js	`88.15% <0.00%> (-7.90%)`	⬇️
lib/core/request.js	`96.93% <0.00%> (-3.07%)`	⬇️
lib/node/http-parser.js

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3d002c1...66af633. Read the comment docs.

ronag · 2021-02-28T15:39:18Z

Things that might be worth moving to native code:

dnlup · 2021-02-28T15:56:21Z

If you have time could you try npm run build on mac? I only have a Linux box.

dnlup · 2021-02-28T15:56:54Z

The llhttp repo is cloned using ssh.

dnlup · 2021-02-28T16:37:26Z

Rebased against #571. I'll fix tests iteratively next.

lib/llhttp/parser.js

dnlup · 2021-03-02T10:38:13Z

I was able to run a benchmark with the current changes, even if not complete:

Node version: 14.16.0

current:

> [email protected] bench /home/dnlup/Dev/dnlup/undici
> npx concurrently -k -s first "node benchmarks/server.js" "node -e 'setTimeout(() => {}, 1000)' && node benchmarks"

[1] http - no agent  x 402 ops/sec ±3.49% (61 runs sampled)
[1] http - keepalive x 453 ops/sec ±2.94% (59 runs sampled)
[1] http - keepalive - multiple sockets x 3,191 ops/sec ±4.19% (72 runs sampled)
[1] undici - pipeline x 3,702 ops/sec ±3.84% (69 runs sampled)
[1] undici - request x 3,572 ops/sec ±3.56% (66 runs sampled)
[1] undici - pool - request - multiple sockets x 3,527 ops/sec ±3.79% (65 runs sampled)
[1] undici - stream x 3,885 ops/sec ±3.37% (71 runs sampled)
[1] undici - dispatch x 3,842 ops/sec ±4.04% (65 runs sampled)
[1] undici - noop x 4,524 ops/sec ±1.35% (67 runs sampled)

llhttp_wasm:

> [email protected] bench /home/dnlup/Dev/dnlup/undici
> npx concurrently -k -s first "node benchmarks/server.js" "node -e 'setTimeout(() => {}, 1000)' && node benchmarks"

[1] http - no agent  x 420 ops/sec ±4.12% (50 runs sampled)
[1] http - keepalive x 441 ops/sec ±2.88% (52 runs sampled)
[1] http - keepalive - multiple sockets x 3,400 ops/sec ±2.81% (78 runs sampled)
[1] undici - pipeline x 3,813 ops/sec ±2.96% (72 runs sampled)
[1] undici - request x 3,578 ops/sec ±2.83% (67 runs sampled)
[1] undici - pool - request - multiple sockets x 3,830 ops/sec ±3.83% (69 runs sampled)
[1] undici - stream x 3,773 ops/sec ±3.81% (65 runs sampled)
[1] undici - dispatch x 3,976 ops/sec ±3.07% (71 runs sampled)
[1] undici - noop x 4,302 ops/sec ±2.46% (58 runs sampled)

Not sure my results are lower than average on both branches, but I thought it would have been worse.

mcollina · 2021-03-02T10:49:08Z

This is incredibly positive. Good work @dnlup!

lib/llhttp/parser.js

dnlup · 2021-03-04T15:20:28Z

The only test remaining is https://github.com/nodejs/undici/blob/master/test/client-pipeline.js#L975 .

The buffers differ by 2 bytes, it looks like the first 2 from the tap output. I manually checked the chunks in the on('data') callback and they have all the same size. The kOnBody is different now because we have direct access to the body buffer and we don't have the offsets from the whole message buffer (at least I could find them, it looks like it's implemented in the Node parser and not in llhttp). I'll see if I can solve this.

ronag · 2021-03-04T16:30:14Z

Could maybe be a string encoding thing?

dnlup · 2021-03-04T17:33:36Z

Could maybe be a string encoding thing?

It makes sense, how could I test it?

build/llhttp.js

dnlup · 2021-03-07T18:15:19Z

I think if there has been an error then the state is pretty much undefined and the parser should not be used anymore. I don't think we do but it might be worth to always fail execute is called after having failed. Maybe unnecessary?

I'll dig into it. There's a reset function in llhttp that we are not using. I'll check what the Node parser is doing.

ronag · 2021-03-07T18:16:30Z

I'll dig into it. There's a reset function in llhttp that we are not using. I'll check what the Node parser is doing.

I don't think we need to fix this since we always destroy the socket and parser on error anyway. I just think we should have an assertion or error as to not make mistakes in the future.

mcollina · 2021-03-07T21:10:38Z

build/llhttp.js

+const { execSync } = require('child_process')
+const TMP = require('os').tmpdir()
+
+const REPO = '[email protected]:dnlup/llhttp.git'


I would prefer if we used the upstream repository instead and we copy over whatever files we need.

Alternatively, I'd be ok in doing that adds this to llhttp itself.

I would prefer if we used the upstream repository instead and we copy over whatever files we need.

Expanding on https://github.com/nodejs/undici/pull/575/files#r589005221 (and thinking out loud), we need to build the c files prior to build the wasm binary (and the source files are the same llhttp is using to build the parser), I think moving that out of llhttp would require us to duplicate a lot of build logic in here. Like, we would need the c src files, the js files to build the parser, add the build scripts.

Alternatively, I'd be ok in doing that adds this to llhttp itself.

I'll open a PR to llhttp and see what happens.

Ok, here's the PR:

nodejs/llhttp#93

This implement a minimal parser that uses llhttp built with wasm. A custom repository was setup to have a custom wasm build. Benchmarks of the current changes: http - no agent x 436 ops/sec ±4.06% (54 runs sampled) http - keepalive x 453 ops/sec ±2.72% (54 runs sampled) http - keepalive - multiple sockets x 3,191 ops/sec ±2.30% (73 runs sampled) undici - pipeline x 4,062 ops/sec ±3.14% (73 runs sampled) undici - request x 3,925 ops/sec ±2.81% (74 runs sampled) undici - pool - request - multiple sockets x 4,045 ops/sec ±3.36% (76 runs sampled) undici - stream x 3,912 ops/sec ±2.69% (74 runs sampled) undici - dispatch x 4,371 ops/sec ±1.96% (79 runs sampled) undici - noop x 4,237 ops/sec ±2.58% (59 runs sampled) Benchmarks of the stable branch: http - no agent x 410 ops/sec ±4.18% (63 runs sampled) http - keepalive x 446 ops/sec ±2.75% (53 runs sampled) http - keepalive - multiple sockets x 3,248 ops/sec ±2.97% (73 runs sampled) undici - pipeline x 3,666 ops/sec ±3.18% (73 runs sampled) undici - request x 3,951 ops/sec ±3.04% (72 runs sampled) undici - pool - request - multiple sockets x 4,017 ops/sec ±3.16% (75 runs sampled) undici - stream x 4,014 ops/sec ±3.89% (70 runs sampled) undici - dispatch x 3,823 ops/sec ±3.46% (67 runs sampled) undici - noop x 4,569 ops/sec ±1.99% (59 runs sampled)

dnlup · 2021-03-08T13:57:41Z

I think if there has been an error then the state is pretty much undefined and the parser should not be used anymore. I don't think we do but it might be worth to always fail execute is called after having failed. Maybe unnecessary?

For what is worth, I think it's already handled inside llhttp by continuing to return the last error if execute returns a non-pause code:

https://github.com/nodejs/llhttp/blob/master/src/native/api.h#L81

That should already cause a request to fail because execute would return an error.

ronag · 2021-03-11T17:34:51Z

Was reading nodejs/node#37676 (comment) and got curious whether our wasm build is strict or loose?

dnlup · 2021-03-11T18:20:51Z

Was reading nodejs/node#37676 (comment) and got curious whether our wasm build is strict or loose?

I suspect it is strict. I don't see much difference between the node build and the wasm build. How should Node disable strict? Maybe removing this definition?

ronag · 2021-03-23T11:29:36Z

@dnlup @mcollina should we try to pull in all the required llhttp changes into undici (i.e. the wasm parts)? Doing things through llhttp seems slow and I feel it will stall progress for future work as well.

dnlup · 2021-03-23T11:50:16Z

@dnlup @mcollina should we try to pull in all the required llhttp changes into undici (i.e. the wasm parts)? Doing things through llhttp seems slow and I feel it will stall progress for future work as well.

We would need to commit the c source, the constants.js file and add some build scripts. These are the first things the come to my mind.

ronag · 2021-03-23T11:55:16Z

I wouldn't mind that.

mcollina · 2021-03-23T12:27:52Z

I'm ok in vendoring the full of the llhttp source and apply patches on top/put the automation in our own tree.

ronag · 2021-03-25T09:08:38Z

@mcollina can we merge this once conflicts are resolved?

mcollina

lgtm

mcollina · 2021-03-25T09:26:47Z

This can land if CI pass

dnlup · 2021-03-25T09:30:03Z

I think CI is fine, it's stuck for some reason. I am waiting on #601 to remove the deps folder from the bundled assets.

dnlup · 2021-03-25T09:31:56Z

Apologies, the comment regards #611 . I think we should merge that PR.

ronag · 2021-03-25T09:32:53Z

Apologies, the comment regards #611 . I think we should merge that PR.

I'm confused. Merge into this PR? Or what?

dnlup · 2021-03-25T09:35:07Z

Apologies, the comment regards #611 . I think we should merge that PR.

I'm confused. Merge into this PR? Or what?

Merge into main that PR directly. It has the same code, it just embeds the native sources and the build script.

ronag · 2021-03-25T09:37:24Z

Merge into main that PR directly. It has the same code, it just embeds the native sources and the build script.

So close this PR?

dnlup · 2021-03-25T09:39:03Z

Merge into main that PR directly. It has the same code, it just embeds the native sources and the build script.

So close this PR?

Yes, you're right. I guess I was waiting to see if this was still doable.

ronag added 5 commits February 27, 2021 01:46

use parser.execute

9bd4b0b

fixup

08ebbed

fixup

723afc5

fixup

aa7531c

fixup

5de2a3c

ronag reviewed Feb 27, 2021

View reviewed changes

lib/llhttp/parser.js Outdated Show resolved Hide resolved

ronag reviewed Feb 27, 2021

View reviewed changes

lib/llhttp/parser.js Outdated Show resolved Hide resolved

ronag added 3 commits February 27, 2021 15:35

fixup

361b560

fixup

2361b8e

Merge branch 'master' into parser-execute

6920db3

mcollina reviewed Feb 28, 2021

View reviewed changes

lib/llhttp/parser.js Show resolved Hide resolved

ronag reviewed Feb 28, 2021

View reviewed changes

lib/llhttp/parser.js Outdated Show resolved Hide resolved

dnlup force-pushed the llhttp_wasm branch from 42f9f09 to 275b204 Compare February 28, 2021 16:35

ronag mentioned this pull request Feb 28, 2021

use parser.execute #571

Closed

ronag reviewed Mar 2, 2021

View reviewed changes

lib/llhttp/parser.js Outdated Show resolved Hide resolved

ronag reviewed Mar 3, 2021

View reviewed changes

lib/llhttp/parser.js Outdated Show resolved Hide resolved

ronag reviewed Mar 7, 2021

View reviewed changes

build/llhttp.js Show resolved Hide resolved

mcollina requested changes Mar 7, 2021

View reviewed changes

dnlup force-pushed the llhttp_wasm branch from 75c565b to 66af633 Compare March 8, 2021 09:58

dnlup mentioned this pull request Mar 8, 2021

Wasm build nodejs/llhttp#93

Merged

Base automatically changed from master to main March 12, 2021 21:33

dnlup mentioned this pull request Mar 23, 2021

Embed llhttp #611

Merged

ronag added this to the 4.0 milestone Mar 24, 2021

ronag assigned dnlup Mar 24, 2021

Merge branch 'main' into llhttp_wasm

e20a2f0

mcollina approved these changes Mar 25, 2021

View reviewed changes

mcollina closed this Mar 25, 2021

puzpuzpuz mentioned this pull request May 18, 2021

Wanted: Reimplement WHATWG URL Parser in WASM nodejs/node#38708

Closed

Use llhttp wasm build #575

Use llhttp wasm build #575

Conversation

dnlup commented Feb 27, 2021 • edited Loading

ronag left a comment

Choose a reason for hiding this comment

mcollina commented Feb 28, 2021

ronag commented Feb 28, 2021 • edited Loading

mcollina commented Feb 28, 2021

codecov-io commented Feb 28, 2021 • edited Loading

Codecov Report

ronag commented Feb 28, 2021

dnlup commented Feb 28, 2021

dnlup commented Feb 28, 2021

dnlup commented Feb 28, 2021

dnlup commented Mar 2, 2021

mcollina commented Mar 2, 2021

dnlup commented Mar 4, 2021

ronag commented Mar 4, 2021 • edited Loading

dnlup commented Mar 4, 2021

dnlup commented Mar 7, 2021

ronag commented Mar 7, 2021 • edited Loading

mcollina Mar 7, 2021

Choose a reason for hiding this comment

dnlup Mar 8, 2021

Choose a reason for hiding this comment

dnlup Mar 8, 2021

Choose a reason for hiding this comment

dnlup commented Mar 8, 2021

ronag commented Mar 11, 2021

dnlup commented Mar 11, 2021

ronag commented Mar 23, 2021 • edited Loading

dnlup commented Mar 23, 2021

ronag commented Mar 23, 2021 • edited Loading

mcollina commented Mar 23, 2021

ronag commented Mar 25, 2021

mcollina left a comment

Choose a reason for hiding this comment

mcollina commented Mar 25, 2021

dnlup commented Mar 25, 2021

dnlup commented Mar 25, 2021

ronag commented Mar 25, 2021

dnlup commented Mar 25, 2021

ronag commented Mar 25, 2021

dnlup commented Mar 25, 2021 • edited Loading

dnlup commented Feb 27, 2021 •

edited

Loading

ronag commented Feb 28, 2021 •

edited

Loading

codecov-io commented Feb 28, 2021 •

edited

Loading

ronag commented Mar 4, 2021 •

edited

Loading

ronag commented Mar 7, 2021 •

edited

Loading

ronag commented Mar 23, 2021 •

edited

Loading

ronag commented Mar 23, 2021 •

edited

Loading

dnlup commented Mar 25, 2021 •

edited

Loading