go analysis memory mallocs

[刘桂祥]

// example1.go
package main


import "runtime"


func main() {
    m := new(runtime.MemStats)
    runtime.ReadMemStats(m)
    println(m.Alloc, m.Mallocs)
    cap := 1024 * 1024 * 3
    s := make([]byte, cap)
    _ = s
    runtime.ReadMemStats(m)
    println(m.Alloc, m.Mallocs)
}

go run example1.go

32744 65

3178472 66

// example2.go
package main


import "runtime"


func main() {
    m := new(runtime.MemStats)
    runtime.ReadMemStats(m)
    println(m.Alloc, m.Mallocs)
    cap := 1024 * 1024 * 4
    s := make([]byte, cap)
    _ = s
    runtime.ReadMemStats(m)
    println(m.Alloc, m.Mallocs)
}

go run example2.go

32744 65

4228984 77

why it have 12 times mallocs ??

[Dave Cheney]

Be careful that the compiler isnt removing some or all of your program. Check the asm to assert that your program is not being optimised away.

Then check -gcflags=-m to see if the compiler is choosing a different escape analysis depending on the size of your allocation.

[刘桂祥]

In my two examples, I use variable in makeslice that it will all escape to the heap ,  But I just want to use runtime.ReadMemStats to lookup it

and find strangely example2.go memstats.Mallocs is many times  don't know why ...

在 2016年9月29日星期四 UTC+8下午2:56:38，Dave Cheney写道：

[Dave Cheney]

One way to do this might be to enable memory profiling in your program with the rate set to 1. Hopefully this will record the stack trace of every allocation. The data may need some interpretation

[刘桂祥]

package main

import "runtime"

func init() {

    runtime.MemProfileRate = 1

}

func main() {

    m := new(runtime.MemStats)

    runtime.ReadMemStats(m)

    println(m.Alloc, m.Mallocs)

    cap := 1024 * 1024 * 4

    s := make([]byte, cap)

    _ = s

    runtime.ReadMemStats(m)

    println(m.Alloc, m.Mallocs)

}

go run example2.go

32744 65

4228984 76

sorry  add this to example2.go  and the results is a little different  but I also don't konw the reason to the results

在 2016年9月29日星期四 UTC+8下午5:07:23，Dave Cheney写道：

[Dave Cheney]

Sorry, you'll have to work harder to generate a memory profile. My github.com/pkg/profile package may help automate some of the work, but may introduce a few allocations of its own which you will have to filter out.

[刘桂祥]

package main

import "runtime"

func main() {

    m := new(runtime.MemStats)

    runtime.ReadMemStats(m)

    cap := 1024

    println("debug0")

    s := make([]byte, cap)

    println("debug1")

    _ = s

    runtime.ReadMemStats(m)

    println("memstats:", m.Alloc, m.Mallocs)

}

I use go1.4.2 and In go source code add mallocgc debug info 

here is the result:

debug0

mallocgc: 1024

mallocgc: 272

mallocgc: 32

debug1

I doubt there is other goroutine call the mallocgc  

 

but if I use go test -bench="." makeslice_test.go

func BenchmarkMakeSlice2(b *testing.B) {

    b.ReportAllocs()

    cap := 1024

    for i := 0; i < b.N; i++ {

        s := make([]byte, cap, cap)

        _ = s

    }

}

the output is:

14120 ns/op    1024 B/op       1 allocs/op

I infer the go test know all the mallocgc and also know which is  user's source code to call it  

my guess

在 2016年9月29日星期四 UTC+8下午5:30:54，Dave Cheney写道：

[Dave Cheney]

You should use Go 1.7.1 or later. 

[刘桂祥]

if use go1.7

in runtime/malloc.go file add mallocgc debug info and go install and go run x.go    I will get a error "signal: segmentation fault"

in go1.4.2 will not

func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {

    println("mallocgc:size=", size, "kind=", typ.kind, "needzero=", needzero)

在 2016年9月29日星期四 UTC+8下午9:04:58，Dave Cheney写道：