Memory overhead of maps in Go

Question

map[byte]byte{0:10} should be using least 2 bytes, one for value and one per key. But as each hashmap implmentation, there is also a hidden cost per item. What is the memory overhead per entry in Go maps in both gccgo and gc?

Answer 1

Here's a cross-platform reimplementation of Nick's program. It includes changes where I think it was flawed. It also adds more measured data points.

Note: To allow for a wider "entries" range, the measured map bellow is map[int16]byte.

package main

import (
        "fmt"
        "runtime"
        "unsafe"
)

func Alloc() uint64 {
        var stats runtime.MemStats
        runtime.GC()
        runtime.ReadMemStats(&stats)
        return stats.Alloc - uint64(unsafe.Sizeof(hs[0]))*uint64(cap(hs))
}

var hs = []*map[int16]byte{}

func main() {
        hs := []*map[int16]byte{}
        n := 1000
        before := Alloc()
        for i := 0; i < n; i++ {
                h := map[int16]byte{}
                hs = append(hs, &h)
        }
        after := Alloc()
        emptyPerMap := float64(after-before) / float64(n)
        fmt.Printf("Bytes used for %d empty maps: %d, bytes/map %.1f\n", n, after-before, emptyPerMap)
        hs = nil

        k := 1
        for p := 1; p < 16; p++ {
                before = Alloc()
                for i := 0; i < n; i++ {
                        h := map[int16]byte{}
                        for j := 0; j < k; j++ {
                                h[int16(j)] = byte(j)
                        }
                        hs = append(hs, &h)
                }
                after = Alloc()
                fullPerMap := float64(after-before) / float64(n)
                fmt.Printf("Bytes used for %d maps with %d entries: %d, bytes/map %.1f\n", n, k, after-before, fullPerMap)
                fmt.Printf("Bytes per entry %.1f\n", (fullPerMap-emptyPerMap)/float64(k))
                k *= 2
        }

}

---

Output

jnml@fsc-r630:~/src/tmp$ go build && ./tmp && go version && uname -a
Bytes used for 1000 empty maps: 146816, bytes/map 146.8
Bytes used for 1000 maps with 1 entries: 147040, bytes/map 147.0
Bytes per entry 0.2
Bytes used for 1000 maps with 2 entries: 147040, bytes/map 147.0
Bytes per entry 0.1
Bytes used for 1000 maps with 4 entries: 247136, bytes/map 247.1
Bytes per entry 25.1
Bytes used for 1000 maps with 8 entries: 439056, bytes/map 439.1
Bytes per entry 36.5
Bytes used for 1000 maps with 16 entries: 818688, bytes/map 818.7
Bytes per entry 42.0
Bytes used for 1000 maps with 32 entries: 1194688, bytes/map 1194.7
Bytes per entry 32.7
Bytes used for 1000 maps with 64 entries: 2102976, bytes/map 2103.0
Bytes per entry 30.6
Bytes used for 1000 maps with 128 entries: 4155072, bytes/map 4155.1
Bytes per entry 31.3
Bytes used for 1000 maps with 256 entries: 6698688, bytes/map 6698.7
Bytes per entry 25.6
Bytes used for 1000 maps with 512 entries: 14142976, bytes/map 14143.0
Bytes per entry 27.3
Bytes used for 1000 maps with 1024 entries: 51349184, bytes/map 51349.2
Bytes per entry 50.0
Bytes used for 1000 maps with 2048 entries: 102467264, bytes/map 102467.3
Bytes per entry 50.0
Bytes used for 1000 maps with 4096 entries: 157214816, bytes/map 157214.8
Bytes per entry 38.3
Bytes used for 1000 maps with 8192 entries: 407031200, bytes/map 407031.2
Bytes per entry 49.7
Bytes used for 1000 maps with 16384 entries: 782616864, bytes/map 782616.9
Bytes per entry 47.8
go version devel +83b0b94af636 Sat Mar 09 16:25:30 2013 +1100 linux/amd64
Linux fsc-r630 3.2.0-38-generic #61-Ubuntu SMP Tue Feb 19 12:18:21 UTC 2013 x86_64 x86_64 x86_64 GNU/Linux
jnml@fsc-r630:~/src/tmp$ 

It's nice to see the numbers are better (by a factor of about 4x). The numbers for the release version (1.0.3) are only slightly higher:

jnml@fsc-r630:~/src/tmp$ go build && ./tmp
Bytes used for 1000 empty maps: 144192, bytes/map 144.2
Bytes used for 1000 maps with 1 entries: 144192, bytes/map 144.2
Bytes per entry 0.0
Bytes used for 1000 maps with 2 entries: 144192, bytes/map 144.2
Bytes per entry 0.0
Bytes used for 1000 maps with 4 entries: 315648, bytes/map 315.6
Bytes per entry 42.9
Bytes used for 1000 maps with 8 entries: 436288, bytes/map 436.3
Bytes per entry 36.5
Bytes used for 1000 maps with 16 entries: 885824, bytes/map 885.8
Bytes per entry 46.4
Bytes used for 1000 maps with 32 entries: 1331264, bytes/map 1331.3
Bytes per entry 37.1
Bytes used for 1000 maps with 64 entries: 2292800, bytes/map 2292.8
Bytes per entry 33.6
Bytes used for 1000 maps with 128 entries: 4935920, bytes/map 4935.9
Bytes per entry 37.4
Bytes used for 1000 maps with 256 entries: 12164160, bytes/map 12164.2
Bytes per entry 47.0
Bytes used for 1000 maps with 512 entries: 29887808, bytes/map 29887.8
Bytes per entry 58.1
Bytes used for 1000 maps with 1024 entries: 56840768, bytes/map 56840.8
Bytes per entry 55.4
Bytes used for 1000 maps with 2048 entries: 108736064, bytes/map 108736.1
Bytes per entry 53.0
Bytes used for 1000 maps with 4096 entries: 184368752, bytes/map 184368.8
Bytes per entry 45.0
Bytes used for 1000 maps with 8192 entries: 431340576, bytes/map 431340.6
Bytes per entry 52.6
Bytes used for 1000 maps with 16384 entries: 815378816, bytes/map 815378.8
Bytes per entry 49.8
jnml@fsc-r630:~/src/tmp$

Answer 2

/*  Hal3 Mon Jul 18 20:58:16 BST 2016 go version go1.5.1 linux/amd64
Bytes used for 1000 empty maps: 0, bytes/map 0.0
Bytes used for 1000 maps with 1 entries: 112192, bytes/map 112.2
Bytes per entry 112.2
Bytes used for 1000 maps with 2 entries: 113472, bytes/map 113.5
Bytes per entry 56.7
Bytes used for 1000 maps with 4 entries: 110912, bytes/map 110.9
Bytes per entry 27.7
Bytes used for 1000 maps with 8 entries: 112192, bytes/map 112.2
Bytes per entry 14.0
Bytes used for 1000 maps with 16 entries: 231600, bytes/map 231.6
Bytes per entry 14.5
Bytes used for 1000 maps with 32 entries: 413768, bytes/map 413.8
Bytes per entry 12.9
Bytes used for 1000 maps with 64 entries: 736920, bytes/map 736.9
Bytes per entry 11.5
Bytes used for 1000 maps with 128 entries: 1419624, bytes/map 1419.6
Bytes per entry 11.1
Bytes used for 1000 maps with 256 entries: 2735192, bytes/map 2735.2
Bytes per entry 10.7
Bytes used for 1000 maps with 512 entries: 5655168, bytes/map 5655.2
Bytes per entry 11.0
Bytes used for 1000 maps with 1024 entries: 10919888, bytes/map 10919.9
Bytes per entry 10.7
Bytes used for 1000 maps with 2048 entries: 21224528, bytes/map 21224.5
Bytes per entry 10.4
Bytes used for 1000 maps with 4096 entries: 42391024, bytes/map 42391.0
Bytes per entry 10.3
Bytes used for 1000 maps with 8192 entries: 84613344, bytes/map 84613.3
Bytes per entry 10.3
Bytes used for 1000 maps with 16384 entries: 169152560, bytes/map 169152.6
Bytes per entry 10.3
Mon Jul 18 20:58:25 BST 2016 */

Answer 3

Overhead per map entry is not a constant value, since it depends on a number of buckets per map entry.

There is a great article on the map internals: https://www.ardanlabs.com/blog/2013/12/macro-view-of-map-internals-in-go.html

The hash table for a Go map is structured as an array of buckets. The number of buckets is always equal to a power of 2.

...

How Maps Grow

As we continue to add or remove key/value pairs from the map, the efficiency of the map lookups begin to deteriorate. The load threshold values that determine when to grow the hash table are based on these four factors:

% overflow : Percentage of buckets which have an overflow bucket

bytes/entry : Number of overhead bytes used per key/value pair

hitprobe : Number of entries that need to be checked when looking up a key

missprobe : Number of entries that need to be checked when looking up an absent key

For example a very simple benchmark can show a dramatic increase in overhead per entry when increasing number of entries just by 1:

func Benchmark(b *testing.B) {
    m := make(map[int64]struct{})

    // also resets mem stats
    b.ResetTimer()

    for i := 0; i < b.N; i++ {
        m[int64(i)] = struct{}{}
    }
}

Benching with 106496 entries:

go test -bench . -benchtime 106496x -benchmem Benchmark-2 106495 65.7 ns/op 31 B/op 0 allocs/op

e.g. 31 bytes per entry

Now increase the number of entries by one:

go test -bench . -benchtime 106497x -benchmem Benchmark-2 106497 65.7 ns/op 57 B/op 0 allocs/op

e.g. 57 bytes per entry

Increasing the number of entries by 1 resulted in the doubling of the number of underlying buckets, which resulted in an additional overhead. The overhead will decrease when more entries are added, until the number of buckets is doubled again.

Answer 4

Here is an experiment to measure the overhead of maps. Works under Linux only.

package main

import (
    "fmt"
    "io/ioutil"
    "log"
    "os"
    "runtime"
    "strconv"
    "strings"
)

func ReadRss() int {
    data, err := ioutil.ReadFile("/proc/self/statm")
    if err != nil {
        log.Fatal(err)
    }
    rss, err := strconv.Atoi(strings.Fields(string(data))[1])
    if err != nil {
        log.Fatal(err)
    }
    return rss * os.Getpagesize()
}

func main() {
    hs := []*map[byte]byte{}
    before := ReadRss()
    n := 10000
    for i := 0; i < n; i++ {
        h := map[byte]byte{}
        hs = append(hs, &h)
    }
    after := ReadRss()
    empty_per_map := float64(after-before)/float64(n)
    fmt.Printf("Bytes used for %d empty maps: %d, bytes/map %.1f\n", n, after-before, empty_per_map)
    hs = nil
    runtime.GC()

    before = ReadRss()
    for i := 0; i < n; i++ {
        h := map[byte]byte{}
        for j := byte(0); j < 100; j++ {
            h[j] = j
        }
        hs = append(hs, &h)
    }
    after = ReadRss()
    full_per_map := float64(after-before)/float64(n)
    fmt.Printf("Bytes used for %d maps with 100 entries: %d, bytes/map %.1f\n", n, after-before, full_per_map)
    fmt.Printf("Bytes per entry %.1f\n", (full_per_map - empty_per_map)/100)

}

It prints this on my 64 bit Linux machine using go 1.0.3

Bytes used for 10000 empty maps: 1695744, bytes/map 169.6
Bytes used for 10000 maps with 100 entries: 43876352, bytes/map 4387.6
Bytes per entry 42.2

Or using go 1.0

Bytes used for 10000 empty maps: 1388544, bytes/map 138.9
Bytes used for 10000 maps with 100 entries: 199323648, bytes/map 19932.4
Bytes per entry 197.9

The memory measurements are done using the Linux OS rather than Go's internal memory stats. The memory measurements are coarse as they are returned in 4k pages, hence the large number of maps created.

So in round numbers each map costs about 170 bytes and each entry costs 42 bytes using go 1.0.3 (much more for 1.0)