随笔- 1793 文章- 1 评论- 61 阅读- 298万

sync.Pool is much slower than using channel, so why should we use sync.Pool?

package client
 
import (
    "runtime"
    "sync"
    "testing"
)
 
type MPool chan interface{}
 
 
type A struct {
    s        string
    b        int
    overflow *[2]*[]*string 
 
}
 
var p = sync.Pool{
    New: func() interface{} { return new(A) },
}
 
var mp MPool = make(chan interface{}, 100)
 
func get() interface{} {
    select {
    case r := <-mp:
        return r
    default:
        return new(A)
    }
}
 
func put(a interface{}) {
    select {
    case mp <- a:
    default:
    }
    return
}
 
func pool() {
    a := p.Get()
    p.Put(a)
}
 
 
func init() {
    runtime.GOMAXPROCS(8)
}
 
func BenchmarkName(b *testing.B) {
    //for i := 0; i < 20; i++ {
    //    p.Put(new(A))
    //}
    //b.ResetTimer()
    //for i := 0; i < b.N; i++ {
    //    for i := 0; i < 100; i++ {
    //        go func() {
    //            p.Put(p.Get())
    //        }()
     //    }
     //}
     for i := 0; i < 20; i++ {
         p.Put(new(A))
     }
     b.ResetTimer()
     for i := 0; i < b.N; i++ {
         for i := 0; i < 100; i++ {
             p.Put(p.Get())
         }
     }
 }
  
 func BenchmarkNotPool(b *testing.B) {
     for i := 0; i < 20; i++ {
         put(new(A))
     }
     b.ResetTimer()
     for i := 0; i < b.N; i++ {
         for i := 0; i < 100; i++ {
             a := get()
             put(a)
         }
     }
 }
                                                            
                                                           

　　测试结果：

原文：https://stackoverflow.com/questions/50851421/sync-pool-is-much-slower-than-using-channel-so-why-should-we-use-sync-pool

-----------------------------------------------------------------------

I read sync.Pool design, but find it is two logic, why we need localPool to solve lock compete. We can just use chan to implement one.

Using channel is 4x times faster than sync.pool!

Besides pool can clear object, what advantage does it have?

This is the pool implementation and benchmarking code:

package client
 
import (
    "runtime"
    "sync"
    "testing"
)
 
type MPool chan interface{}
 
 
type A struct {
    s        string
    b        int
    overflow *[2]*[]*string 
 
}
 
var p = sync.Pool{
    New: func() interface{} { return new(A) },
}
 
var mp MPool = make(chan interface{}, 100)
 
func get() interface{} {
    select {
    case r := <-mp:
        return r
    default:
        return new(A)
    }
}
 
func put(a interface{}) {
    select {
    case mp <- a:
    default:
    }
    return
}
 
func pool() {
    a := p.Get()
    p.Put(a)
}
 
 
func init() {
    runtime.GOMAXPROCS(8)
}
 
func BenchmarkName(b *testing.B) {
    for i := 0; i < 20; i++ {
        p.Put(new(A))
    }
    b.ResetTimer()
    for i := 0; i < b.N; i++ {
        for i := 0; i < 100; i++ {
            go func() {
                p.Put(p.Get())
            }()
        }
    }
}
 
func BenchmarkNotPool(b *testing.B) {
    for i := 0; i < 20; i++ {
        put(new(A))
    }
    b.ResetTimer()
    for i := 0; i < b.N; i++ {
        for i := 0; i < 100; i++ {
            a := get()
            put(a)
        }
    }
}

　　Answer:

You are not benchmarking the same thing, so you can't compare the results.

BenchmarkName() launches goroutines which have significant overheard and you don't even wait for those goroutines to finish, while BenchmarkNotPool() just gets and puts an object in the pool in the same goroutine.

If you modify BenchmarkName() to do the same, the benchmark results actually show it's the other way: sync.Pool is more than 3 times faster, which is true, so that's its use / advantage.

func BenchmarkName(b *testing.B) {
    for i := 0; i < 20; i++ {
        p.Put(new(A))
    }
    b.ResetTimer()
    for i := 0; i < b.N; i++ {
        for i := 0; i < 100; i++ {
            p.Put(p.Get())
        }
    }
}

Results:

BenchmarkName-8           500000              2453 ns/op
BenchmarkNotPool-8        200000              7984 ns/op

Also see related question: How to implement Memory Pooling in Golang

Thanks for your great answer, my bad, forgot to do go in BechmarkNotPool, but in adverse we proof use sync.Pool is better.

– pete lin

Jun 14 '18 at 7:49

@petelin You can't just launch goroutines in the middle of the benchmark function, and not wait for them to complete. That will render the benchmark result completely useless.
– icza
Jun 14 '18 at 8:09