A very fast arena allocator for Go

This package provides a very fast almost lock-free arena allocator for Go. "Almost" lock-free because while individual allocations are a single atomic add, a lock is acquired when expanding the arena after a chunk has been exhausted.

Warning

Note that continuing to use any existing data allocated from the arena after a Reset() will result in undefined behaviour.

Usage

alloc := arena.Create(32 * 1024 * 1024) // 32MB chunk size
m := arena.New[MyStruct](alloc)
s := arena.Slice[MyStruct](alloc, 0, 100)
s = arena.Append(alloc, s, MyStruct{})

// Once done, reset the arena to reuse the memory.
alloc.Reset()

Performance

For allocating individual objects, the arena allocator is about an order of magnitude faster than the Go runtime and, of course, has zero allocations and minimal GC overhead.

~/dev/arena $ go test -benchmem -bench .
goos: darwin
goarch: arm64
pkg: github.com/alecthomas/arena
cpu: Apple M3
BenchmarkArena/100-8               6155918         197 ns/op         0 B/op         0 allocs/op
BenchmarkArena/1000-8               599896        1976 ns/op         0 B/op         0 allocs/op
BenchmarkArena/10000-8               61201       19957 ns/op         0 B/op         0 allocs/op
BenchmarkArena/100000-8               6133      198741 ns/op         0 B/op         0 allocs/op
BenchmarkArena/1000000-8               588     2025277 ns/op         0 B/op         0 allocs/op
BenchmarkGoRuntime/100-8            641896        1809 ns/op      8000 B/op       100 allocs/op
BenchmarkGoRuntime/1000-8            66020       18014 ns/op     80000 B/op      1000 allocs/op
BenchmarkGoRuntime/10000-8            5764      209803 ns/op    800011 B/op     10000 allocs/op
BenchmarkGoRuntime/100000-8            560     2114053 ns/op   8000037 B/op    100000 allocs/op
BenchmarkGoRuntime/1000000-8            56    21004230 ns/op  80000029 B/op   1000000 allocs/op
PASS
ok    github.com/alecthomas/arena  14.069s