Go 1.20: memory arenas

Go 1.20 release added a new experimental arena package that provides memory arenas. This article explains what memory arenas are and how you can use them to reduce garbage collection overhead and make your programs faster.

Garbage collection overhead

Go is a garbage-collected language and so it can automatically free allocated objects for you. Go runtime achieves that by periodically running a garbage-collection algorithm that frees unreachable objects. Such automatic memory management simplifies the writing of Go applications and ensures memory safety.

However, large Go programs have to spend a significant amount of CPU time doing garbage collection. In addition, the memory usage is often larger than necessary, because Go runtime delays garbage collection as long as possible to free more memory in a single run.

Memory arenas

Memory arenas allow to allocate objects from a contiguous region of memory and free them all at once with minimal memory management or garbage collection overhead.

You can use memory arenas in functions that allocate a large number of objects, process them for a while, and then free all of the objects at the end.

Memory arenas is an experimental feature available in Go 1.20 behind the GOEXPERIMENT=arenas environment variable:

GOEXPERIMENT=arenas go run main.go

The API and implementation of memory arenas is completely unsupported and Go team makes no guarantees about compatibility or whether it will even continue to exist in any future release.

For example:

import "arena"

type T struct{
    Foo string
    Bar [16]byte
}

func processRequest(req *http.Request) {
    // Create an arena in the beginning of the function.
    mem := arena.NewArena()
    // Free the arena in the end.
    defer mem.Free()

    // Allocate a bunch of objects from the arena.
    for i := 0; i < 10; i++ {
        obj := arena.New[T](mem)
    }

    // Or a slice with length and capacity.
    slice := arena.MakeSlice[T](mem, 100, 200)
}

If you want to use the object allocated from an arena after the arena is freed, you can Clone the object to get a shallow copy allocated from the heap:

mem := arena.NewArena()

obj1 := arena.New[T](mem) // arena-allocated
obj2 := arena.Clone(obj1) // heap-allocated
fmt.Println(obj2 == obj1) // false

mem.Free()

// obj2 can be safely used here

You can also use memory arenas with the reflect package:

var typ = reflect.TypeOf((*T)(nil)).Elem()

mem := arena.NewArena()
defer mem.Free()

value := reflect.ArenaNew(mem, typ)
fmt.Println(value.Interface().(*T))

Address sanitizer

To detect invalid usage patterns, you can use memory arenas with the address sanitizer (asan) and the memory sanitizer (msan).

For example, the following program uses the object after the arena is freed:

package main

import (
    "arena"
)

type T struct {
    Num int
}

func main() {
    mem := arena.NewArena()
    o := arena.New[T](mem)
    mem.Free()
    o.Num = 123 // incorrect: use after free
}

You can run the program with the address sanitizer to get a detailed error message:

go run -asan main.go

accessed data from freed user arena 0x40c0007ff7f8
fatal error: fault
[signal SIGSEGV: segmentation violation code=0x2 addr=0x40c0007ff7f8 pc=0x4603d9]

goroutine 1 [running]:
runtime.throw({0x471778?, 0x404699?})
    /go/src/runtime/panic.go:1047 +0x5d fp=0x10c000067ef0 sp=0x10c000067ec0 pc=0x43193d
runtime.sigpanic()
    /go/src/runtime/signal_unix.go:851 +0x28a fp=0x10c000067f50 sp=0x10c000067ef0 pc=0x445b8a
main.main()
    /workspace/main.go:15 +0x79 fp=0x10c000067f80 sp=0x10c000067f50 pc=0x4603d9
runtime.main()
    /go/src/runtime/proc.go:250 +0x207 fp=0x10c000067fe0 sp=0x10c000067f80 pc=0x434227
runtime.goexit()
    /go/src/runtime/asm_amd64.s:1598 +0x1 fp=0x10c000067fe8 sp=0x10c000067fe0 pc=0x45c5a1

Slices

You can allocate slices using MakeSlice method:

// Alloc []string
slice := arena.MakeSlice[string](mem, length, capacity)

If a slice must be grown to accommodate new elements, you need to allocate a new slice or the slice will be moved to the heap when growing with append:

slice := arena.MakeSlice[string](mem, 0, 0) // empty slice from the arena
slice = append(slice, "") // the slice is on the heap now

You might also consider using other data structures instead of slices, for example, a linked list can be grown without an issue.

Maps

Currently, Go arenas don't support maps, but you can create a user-defined generic map that allows optionally specifying an arena for use in allocating new elements.

Beware of the string

Go memory arenas don't allow to allocate strings directly, but you can get around by allocating a []byte and using unsafe:

src := "source string"

mem := arena.NewArena()
defer mem.Free()

bs := arena.MakeSlice[byte](mem, len(src), len(src))
copy(bs, src)
str := unsafe.String(&bs[0], len(bs))

Such arena-allocated strings can't be used after you free the arena, so be careful when allocating strings from arenas and use address sanitizer

Nil arenas

Nil arenas are not valid, for example, you can't do this to allocate from the heap when using arenas:

obj := arena.New[Object](nil)

You also can't have an Allocator interface, because arena.New is a package method. As a result, code-paths for arena/non-arena code must be separate.

Performance

By using memory arenas, Google has achieved savings of up to 15% in CPU and memory usage for a number of large applications, mainly due to reduction in garbage collection CPU time and heap memory usage.

You can achieve even better results in small toy applications. For example, you can take the Binary Trees example from Benchmark Games and change the code to use memory arenas:

+// Allocate an empty tree node, using an arena if provided.
+func allocTreeNode(a *arena.Arena) *Tree {
+       if a != nil {
+               return arena.New[Tree](a)
        } else {
                return &Tree{}
        }
 }

Then compare the performance of the code without memory arenas:

/usr/bin/time go run arena_off.go
77.27user 1.28system 0:07.84elapsed 1001%CPU (0avgtext+0avgdata 532156maxresident)k
30064inputs+2728outputs (551major+292838minor)pagefaults 0swaps

With the code that uses memory arenas:

GOEXPERIMENT=arenas /usr/bin/time go run arena_on.go
35.25user 5.71system 0:05.09elapsed 803%CPU (0avgtext+0avgdata 385424maxresident)k
48inputs+3320outputs (417major+63931minor)pagefaults 0swaps

The code that uses memory arenas not only runs faster, but also uses less memory:

Metric	Without arenas	With areans
User	77.27	35.25
System	1.28	5.71
Elapsed	0:07.84	0:05.09
RSS	532156k	385424k

Memory arenas in Uptrace

Uptrace is an open source APM tool written in Go. You can use it to monitor applications and set up alerts to receive notifications via email, Slack, Telegram, and more.

Since Uptrace receives data from OpenTelemetry in large batches (1k-10k items), it could use memory arenas to allocate large number of spans and metrics without involving garbage collector.

Another allocation-heavy part of the app that could benefit from memory arenas is Protobuf decoding, which also decodes spans and metrics in large batches.