从入门到掉坑：Go 内存池/对象池技术介绍

package main

import (
  "fmt"
  "os"
  "runtime"
  "time"
)

// Results of this program on my machine:
//
// for t in 1 2 3 4 5; do go run maps.go $t; done
//
// Higher parallelism does help, to some extent:
//
// for t in 1 2 3 4 5; do GOMAXPROCS=8 go run maps.go $t; done
//
// Output(go 1.14):
// With map[int32]*int32, GC took 456.159324ms
// With map[int32]int32, GC took 10.644116ms
// With map shards ([]map[int32]*int32), GC took 383.296446ms
// With map shards ([]map[int32]int32), GC took 1.023655ms
// With a plain slice ([]main.t), GC took 172.776µs

func main() {
  const N = 5e7 // 5000w

  if len(os.Args) != 2 {
    fmt.Printf("usage: %s [1 2 3 4]\n(number selects the test)\n", os.Args[0])
    return
  }

  switch os.Args[1] {
  case "1":
    // Big map with a pointer in the value
    m := make(map[int32]*int32)
    for i := 0; i < N; i++ {
      n := int32(i)
      m[n] = &n
    }
    runtime.GC()
    fmt.Printf("With %T, GC took %s\n", m, timeGC())
    _ = m[0] // Preserve m until here, hopefully
  case "2":
    // Big map, no pointer in the value
    m := make(map[int32]int32)
    for i := 0; i < N; i++ {
      n := int32(i)
      m[n] = n
    }
    runtime.GC()
    fmt.Printf("With %T, GC took %s\n", m, timeGC())
    _ = m[0]
  case "3":
    // Split the map into 100 shards
    shards := make([]map[int32]*int32, 100)
    for i := range shards {
      shards[i] = make(map[int32]*int32)
    }
    for i := 0; i < N; i++ {
      n := int32(i)
      shards[i%100][n] = &n
    }
    runtime.GC()
    fmt.Printf("With map shards (%T), GC took %s\n", shards, timeGC())
    _ = shards[0][0]
  case "4":
    // Split the map into 100 shards
    shards := make([]map[int32]int32, 100)
    for i := range shards {
      shards[i] = make(map[int32]int32)
    }
    for i := 0; i < N; i++ {
      n := int32(i)
      shards[i%100][n] = n
    }
    runtime.GC()
    fmt.Printf("With map shards (%T), GC took %s\n", shards, timeGC())
    _ = shards[0][0]
  case "5":
    // A slice, just for comparison to show that
    // merely holding onto millions of int32s is fine
    // if they're in a slice.
    type t struct {
      p, q int32
    }
    var s []t
    for i := 0; i < N; i++ {
      n := int32(i)
      s = append(s, t{n, n})
    }
    runtime.GC()
    fmt.Printf("With a plain slice (%T), GC took %s\n", s, timeGC())
    _ = s[0]
  }
}

func timeGC() time.Duration {
  start := time.Now()
  runtime.GC()
  return time.Since(start)
}

[1] With map[int32]*int32, GC took 456.159324ms
[2] With map[int32]int32, GC took 10.644116ms
[3] With map shards ([]map[int32]*int32), GC took 383.296446ms
[4] With map shards ([]map[int32]int32), GC took 1.023655ms
[5] With a plain slice ([]main.t), GC took 172.776µs

// With map[int32]int32, GC took 11.285541ms
gc 1 @0.001s 7%: 0.010+2.1+0.012 ms clock, 0.12+0.99/2.1/1.2+0.15 ms cpu, 4->6->6 MB, 5 MB goal, 12 P
...
gc 8 @2.374s 0%: 0.003+3.9+0.018 ms clock, 0.042+0.31/6.7/3.1+0.21 ms cpu, 649->649->537 MB, 650 MB goal, 12 P
gc 9 @4.834s 0%: 0.003+7.5+0.021 ms clock, 0.040+0/14/5.1+0.25 ms cpu, 1298->1298->1073 MB, 1299 MB goal, 12 P
gc 10 @9.188s 0%: 0.003+26+0.004 ms clock, 0.045+0/26/0.35+0.053 ms cpu, 1183->1183->746 MB, 2147 MB goal, 12 P (forced)
gc 11 @9.221s 0%: 0.018+9.4+0.003 ms clock, 0.22+0/17/5.0+0.043 ms cpu, 746->746->746 MB, 1492 MB goal, 12 P (forced)

// With map shards ([]map[int32]int32), GC took 1.017494ms
gc 1 @0.001s 7%: 0.010+2.9+0.048 ms clock, 0.12+0.26/3.6/4.1+0.57 ms cpu, 4->7->6 MB, 5 MB goal, 12 P
...
gc 12 @3.924s 0%: 0.003+3.2+0.004 ms clock, 0.040+1.2/7.5/14+0.048 ms cpu, 822->827->658 MB, 840 MB goal, 12 P
gc 13 @8.096s 0%: 0.003+6.1+0.004 ms clock, 0.044+6.0/14/32+0.053 ms cpu, 1290->1290->945 MB, 1317 MB goal, 12 P
gc 14 @11.619s 0%: 0.003+1.2+0.004 ms clock, 0.045+0/2.5/3.7+0.056 ms cpu, 1684->1684->1064 MB, 1891 MB goal, 12 P (forced)
gc 15 @11.628s 0%: 0.003+0.91+0.004 ms clock, 0.038+0/2.3/3.6+0.057 ms cpu, 1064->1064->1064 MB, 2128 MB goal, 12 P (forced)

// Cache is an LRU cache. It is not safe for concurrent access.
type Cache struct {
  cache map[interface{}]*list.Element
}

type cacheShard struct {
  hashmap     map[uint64]uint32        // key在entries中的位置
  entries     queue.BytesQueue         // 实际是[]byte，新数据来了后copy到尾部
}

redis     [  0.1] 0.00%
redis     [  0.5] 0.38%
redis     [    1] 3.48%
redis     [    5] 71.94%
redis     [   10] 22.90%
redis     [   20] 1.21%
redis     [   40] 0.07%
redis     [ +Inf] 0.01%

bigcache  [  0.1] 0.40%
bigcache  [  0.5] 16.16%
bigcache  [    1] 14.82%
bigcache  [    5] 53.07%
bigcache  [   10] 14.85%
bigcache  [   20] 0.66%
bigcache  [   40] 0.03%
bigcache  [ +Inf] 0.01%

groupcache[  0.1] 0.24%
groupcache[  0.5] 9.59%
groupcache[    1] 9.69%
groupcache[    5] 58.74%
groupcache[   10] 19.10%
groupcache[   20] 2.45%
groupcache[   40] 0.17%
groupcache[ +Inf] 0.03%

goos: linux
goarch: amd64
pkg: open-wuji/go-sdk/wujiclient
BenchmarkClient_Get-8              1000000        1154 ns/op           1.00 hits        87 B/op        3 allocs/op
BenchmarkFilter_Get-8              4899364         302 ns/op           1.00 hits         7 B/op        1 allocs/op
BenchmarkClient_GetParallel-8      8383149         162 ns/op           1.00 hits        80 B/op        2 allocs/op
BenchmarkFilter_GetParallel-8     13053680        91.4 ns/op           1.00 hits         0 B/op        0 allocs/op
PASS
ok    open-wuji/go-sdk/wujiclient 93.494s
Success: Benchmarks passed.