Beaver

Go memory allocators. No CGO.

Inspired by mimalloc's slab and lock-free design, but implemented without C calls.

---

Packages

Package	Description
alloc	Hybrid allocator (auto small/large split), HTTP middleware, JSON helpers
pure	make([]byte) bump allocator with atomic offset
balloc	mmap-based off-heap block allocator
arena	GC-friendly reference-counted allocator

---

Quick Start

package main

import (
    "net/http"
    "github.com/gosuda/beaver/alloc"
)

func main() {
    pool := alloc.NewPool(alloc.HybridFactory(64 << 20))

    mux := http.NewServeMux()
    mux.HandleFunc("/data", func(w http.ResponseWriter, r *http.Request) {
        ctx := r.Context()
        rows, _ := alloc.MakeSlice[Row](ctx, 10000, 10000)
        _ = rows
        w.WriteHeader(http.StatusOK)
    })

    http.ListenAndServe(":8080", alloc.Middleware(pool)(mux))
}

HybridFactory automatically selects:

• <= 4KB: pure Go slab (atomic.Int64 bump)
• > 4KB: mmap off-heap

---

Benchmarks

Same machine: AMD Ryzen 5 5600X, Go 1.24.4, Linux.

Scenario	unsafe-risk/mi (CGO)	Beaver Hybrid	Go Heap
Small 1KB	16,836 ns/op	324 ns/op	478 ns/op
Buffer 4KB	241 ns/op	17 ns/op	485 ns/op
HTTP 32KB	15,564 ns/op	22,850 ns/op	14,220 ns/op
JSON Marshal	17,164 ns/op	16,931 ns/op	16,205 ns/op
API Gateway	361,804 ns/op	438,907 ns/op	413,153 ns/op

See docs/benchmarks.md for details.

---

Documentation

• docs/quickstart.md — Usage examples
• docs/architecture.md — Design notes
• docs/benchmarks.md — Measurement methodology and results
• docs/p99-latency.md — Tail latency tests
• docs/webapp-guide.md — Integration patterns
• docs/api-reference/ — Package API docs

---

License

MIT