性能疑问

[lijialiang]

在本地测试 Small input 中 javascript 性能比 napi 高效

napi @ Small input x 1,431,991 ops/sec ±0.91% (84 runs sampled)
napi#buff @ Small input x 1,535,086 ops/sec ±0.50% (92 runs sampled)
napi#asyncBuff @ Small input x 102,749 ops/sec ±0.46% (82 runs sampled)
javascript @ Small input x 2,553,005 ops/sec ±0.10% (99 runs sampled)
Small input bench suite: Fastest is javascript @ Small input

• node: v14.15.1
• https://github.com/lijialiang/fast-escape

[lijialiang]

原因来自于注释了

#[global_allocator]
static ALLOC: mimalloc::MiMalloc = mimalloc::MiMalloc;

重新测试后的结果如下

napi @ Small input x 2,459,825 ops/sec ±0.38% (89 runs sampled)
napi#buff @ Small input x 2,947,302 ops/sec ±0.15% (93 runs sampled)
napi#asyncBuff @ Small input x 108,739 ops/sec ±0.49% (83 runs sampled)
javascript @ Small input x 2,554,025 ops/sec ±0.13% (99 runs sampled)
Small input bench suite: Fastest is napi#buff @ Small input

[Brooooooklyn]

allocator 会极大程度的影响对内存频繁申请释放的程序性能

[lijialiang]

对于我另外一个案例似乎不起作用，且目前性能差异相距甚大，请教一下可以有哪些具体的优化方向？（例如：simd）

正则转换 html 内容类似 <div>{{ data }}</div> 编译为 <div>hello</div>

js

exports.compileHtmlStr = (html, data) => {
  const reg = new RegExp(/{{\s*data\s*}}/g)
  return html.replace(reg, data)
}

napi + rust

#[macro_use]
extern crate napi_derive;

use regex::Regex;
use napi::{JsObject, JsString, Result, CallContext};

#[global_allocator]
static ALLOC: mimalloc::MiMalloc = mimalloc::MiMalloc;

#[module_exports]
fn init(mut exports: JsObject) -> Result<()> {
  exports.create_named_method("compileHtmlStr", compile_html_str)?;
  Ok(())
}

#[js_function(2)]
fn compile_html_str(ctx: CallContext) -> Result<JsString> {
  let html = ctx.get::<JsString>(0)?.into_utf8()?;
  let data = ctx.get::<JsString>(1)?.into_utf8()?;
  let reg = Regex::new(r"\{\{\s*data\s*\}\}").unwrap();
  let c_html = reg.replace_all(html.as_str()?, data.as_str()?);
  ctx.env.create_string_from_std(c_html.to_string())
}

测试 bench 结果

CompileHtmlStr#Node x 5,543,011 ops/sec ±1.16% (88 runs sampled)
CompileHtmlStr#Rust x 43,318 ops/sec ±0.47% (95 runs sampled)
Fastest is CompileHtmlStr#Node

[Brooooooklyn]

regex 这个包的 SIMD 加速效果不如 escape 这个场景这么明显，无法抵消 Node-API call overhead 的消耗