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

Question

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)
        }
    }
}

Mercie

Answer 1

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

Boot

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

@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. – Boot

@petelin If you had also launched go routines in BechmarkNotPool, that benchmark would also been worthless. In both these cases you should NOT launch go routines. And if you do launch go routines in your benchmark, you need to write the code so you also measure the time those go routines spend, instead of just launching them and let them run "in the background" – Sophomore

Note that it is likely even faster in go1.13 thanks to removing the lock in sync.Pool my findings are 4x faster rather than the 3x stated in the answer medium.com/a-journey-with-go/… – Mada