WebAssembly(WASM)性能逊于Shell？求对比WAT文件及解惑

为什么你的WASM程序更慢？ #

• 瓶颈不在计算，在IO：你测试的核心操作是循环打印1000万条数据，这个场景下IO开销占了90%以上，WASM的计算优势根本发挥不出来。反而sh脚本依赖系统原生的输出逻辑，和系统IO的交互更直接高效；而emcc编译的程序默认会通过emscripten的封装层转发输出，额外的调用链路带来了不小的性能损耗。
• 编译优化没开启：emcc默认是Debug编译模式，没有做任何性能优化。如果加上-O2或-O3参数编译，WASM的执行效率会大幅提升——虽然IO瓶颈还是存在，但计算部分的开销会被压到最低。
• WASM的性能优势场景用错了：WASM擅长的是CPU密集型计算（比如复杂算法、大数据处理），不是频繁读写终端/文件的IO密集型任务。只有当程序核心是纯计算时，WASM的性能才会接近甚至超过原生程序。

Linux CLI可用的WAT性能对比示例 #

下面是一个计算1到1亿整数和的WAT程序（保存为sum.wat），可以用wasmtime或wasmer直接运行，这个场景能体现WASM的计算性能优势：

(module
  (func $main (export "_start")
    (local $i i64)
    (local $sum i64)
    (local.set $i (i64.const 1))
    (local.set $sum (i64.const 0))
    (loop $loop
      (local.set $sum (i64.add (local.get $sum) (local.get $i)))
      (local.set $i (i64.add (local.get $i) (i64.const 1)))
      (br_if $loop (i64.lt_s (local.get $i) (i64.const 100000001)))
    )
    (call $print_i64 (local.get $sum))
  )

  (func $print_i64 (param $n i64)
    (local $buf i32)
    (local $len i32)
    (local $ptr i32)
    (local.set $buf (call $malloc (i32.const 20)))
    (local.set $ptr (local.get $buf))

    (if (i64.eq (local.get $n) (i64.const 0))
      (then
        (i32.store8 (local.get $ptr) (i32.const 48))
        (local.set $len (i32.const 1))
      )
      (else
        (local $temp i64)
        (local.set $temp (local.get $n))
        (local.set $len (i32.const 0))
        (loop $reverse
          (i32.store8 (local.get $ptr) (i32.add (i32.const 48) (i32.wrap_i64 (i64.rem_u (local.get $temp) (i64.const 10)))))
          (local.set $ptr (i32.add (local.get $ptr) (i32.const 1)))
          (local.set $len (i32.add (local.get $len) (i32.const 1)))
          (local.set $temp (i64.div_u (local.get $temp) (i64.const 10)))
          (br_if $reverse (i64.ne (local.get $temp) (i64.const 0)))
        )
        (local $start i32)
        (local $end i32)
        (local.set $start (local.get $buf))
        (local.set $end (i32.sub (i32.add (local.get $buf) (local.get $len)) (i32.const 1)))
        (loop $swap
          (if (i32.lt (local.get $start) (local.get $end))
            (then
              (local $temp_byte i32)
              (local.set $temp_byte (i32.load8_u (local.get $start)))
              (i32.store8 (local.get $start) (i32.load8_u (local.get $end)))
              (i32.store8 (local.get $end) (local.get $temp_byte))
              (local.set $start (i32.add (local.get $start) (i32.const 1)))
              (local.set $end (i32.sub (local.get $end) (i32.const 1)))
              (br $swap)
            )
          )
        )
      )
    )

    (call $fd_write
      (i32.const 1)
      (local.get $buf)
      (i32.const 1)
      (i32.const 0)
    )
    (call $free (local.get $buf))

    (local.set $buf (call $malloc (i32.const 1)))
    (i32.store8 (local.get $buf) (i32.const 10))
    (call $fd_write
      (i32.const 1)
      (local.get $buf)
      (i32.const 1)
      (i32.const 0)
    )
    (call $free (local.get $buf))
  )

  (import "wasi_snapshot_preview1" "malloc" (func $malloc (param i32) (result i32)))
  (import "wasi_snapshot_preview1" "free" (func $free (param i32)))
  (import "wasi_snapshot_preview1" "fd_write" (func $fd_write (param i32 i32 i32 i32) (result i32)))
  (memory 1)
)

运行步骤： #

1. 安装wasmtime：Linux下用包管理器，比如apt install wasmtime或dnf install wasmtime
2. 将代码保存为sum.wat
3. 编译为WASM：wasmtime compile sum.wat -o sum.wasm
4. 运行并计时：time wasmtime sum.wasm

你可以对比用C写的同功能程序（编译时加-O3）和sh脚本（用循环计算和会慢很多），就能直观看到WASM在计算密集场景下的性能优势。

内容的提问来源于stack exchange，提问作者aze2201