Performance bottlenecks and findings

[Boshen]

A summary of perf bottlenecks and findings.

General background research study:

follow the work of https://github.com/nnethercote, read his blog, read the perf book

[Boshen]

Summary	Issue / PR	Prod	Dev
Use precomputed hash in HashMap<Ustr, _>	#1567 #1568	6%	3%
Remove extra source code clone in javascript plugin	#1571 #1579	12%	2%
Enable concurrent mode in swc minifier	#1583	13%	-
Use rustc_hash::FxHasher	#1589 #1595	5%	5%
cache module identifier for NormalModule	#1636	5%	2%
use String::from_utf8_unchecked for codegen	#1648	5%	5%
use BufWriter for file write	#1657	2%	2%

[Boshen]

nodejs_resolver

See the list of PRs

Overall we (with @bvanjoi) have reduced the application time by a constant of 1 to 2 second, and ~10% for large applications.

• Reduce syscall
  • Removed std::fs::canonicalize, which is an expensive syscall.
  • Reduced std::fs::metadata stat call by half
  • Reduced cpu time of doing syscalls by ~15 times
  • Removed useless stat calls on package.json, e.g. test.js/package.json
• Reduce memory usage
  • Removed all unnecessary PathBuf allocations
• Reduce CPU usage
  • Removed unnecessary usage of SmolStr
  • Added rustc_hash::FxHasher for faster hashing

[Boshen]

When writing file to disk, the string is currently joined together all at once, which creates a huge memory allocation pressure.

https://github.com/modern-js-dev/rspack/blob/45c912cc0c5a3b5aa2d9fa5c721638c6ae647a4b/crates/rspack_core/src/compiler/mod.rs#L205

https://github.com/modern-js-dev/rspack-sources/blob/1969fb3d48fdab5bbb4d0a3b81a8aa87e1867954/src/concat_source.rs#L85-L93

Instead, we should just iterate the sources and write it into a BufferWriter to balance the performance around memory allocation and the fs::write syscall.

[Boshen]

Memory Footprint

4.5kiB

Clone source before parse

rspack/crates/rspack_plugin_javascript/src/plugin.rs

Lines 260 to 265 in 20f0fd9

	let (mut ast, diagnostics) = match crate::ast::parse(
	source.source().to_string(),
	syntax,
	&resource_data.resource_path.to_string_lossy(),
	module_type,
	) {

Why do we need to clone the whole source before parsing? It should be able to take a reference instead.

2.5KiB

This is really heavy:

112 Bytes

There's a very frequent allocation here somewhere, but I can't find it.

18.86 MiB

---

It feels like AST is not dropped at all.

[hyf0]

I look into the SWC.

It looks like SWC requires mutable String to remove possible bom.

https://github.com/swc-project/swc/blob/66247f0c49e890aca94d1788be41e1d36d21ed2f/crates/swc_common/src/syntax_pos.rs#L879

[kdy1]

I made https://github.com/swc-project/swc/pull/7029/files
You can remove bom earlier and store it as Arc<String>

[Boshen]

SWC

SWC AST is not fine tuned, for example:

Expression

Is 112 bytes, which puts a lot of allocation pressure all over the place.

Some of the enums can be shrink by boxing their attributes, it would be nice if we can get Expr below 64 bytes. This allows applies for Stmt.

Reallocate in parse_block_body

https://github.com/swc-project/swc/blob/f32be0e7c60ade1288e27731996de428b957dff8/crates/swc_ecma_parser/src/parser/stmt.rs#L72

https://github.com/swc-project/swc/blob/f32be0e7c60ade1288e27731996de428b957dff8/crates/swc_ecma_parser/src/parser/stmt.rs#L29

https://github.com/thomcc/rust-typed-arena/blob/a6abfd34491ba7bdb72255c853ba6fdccffac112/src/lib.rs#L433-L445

Allocate all statements

https://github.com/swc-project/swc/blob/f32be0e7c60ade1288e27731996de428b957dff8/crates/swc_ecma_transforms_optimization/src/simplify/branch/mod.rs#L1293

Short lived 16 bytes in read_number

[Boshen]

I wrote an issue to swc swc-project/swc#7019

[kdy1]

The parse_block_body issue will be fixed if we box more statements

[kdy1]

Actually I considered it in the past, but didn't do it because I have other tasks to work on

[Boshen]

CPU

This is blocking everything for 500ms.

[hyf0]

It should be improved after merging swc-project/swc#7030.

[Boshen]

CPU Utilization