Merge pull request #946 from devlights/add-sliceclone-benchmark

Add benchmarks/03.slice_clone

Author: devlights

examples/benchmarks/03.slice_clone/Taskfile.yml

@@ -0,0 +1,10 @@
+# https://taskfile.dev
+
+version: '3'
+
+tasks:
+  default:
+    cmds:
+      - go test -bench . -benchmem -count 3 | tee bench.txt
+      - defer: rm -f bench.txt
+      - benchstat bench.txt

examples/benchmarks/03.slice_clone/slice_clone_test.go

@@ -0,0 +1,174 @@
+package main
+
+import (
+	"slices"
+	"testing"
+)
+
+const (
+	NumItems = 5 * 1024 * 1024
+)
+
+// コンパイラの最適化を防ぐためのダミー関数
+func voidfn(_ []int) {}
+
+// 方法1: 事前に容量を確保してからappendする方法
+func BenchmarkPreAllocatedAppend(b *testing.B) {
+	var (
+		items = make([]int, 0, NumItems)
+	)
+	// テストデータの準備
+	for i := range NumItems {
+		items = append(items, i)
+	}
+
+	for b.Loop() {
+		var (
+			dst = make([]int, 0, NumItems)
+		)
+		// appendを使用してコピー
+		dst = append(dst, items...)
+		voidfn(dst) // コンパイラの最適化を防ぐ
+	}
+}
+
+// 方法2: 空スライスに直接appendする方法
+func BenchmarkEmptySliceAppend(b *testing.B) {
+	var (
+		items = make([]int, 0, NumItems)
+	)
+	for i := range NumItems {
+		items = append(items, i)
+	}
+
+	for b.Loop() {
+		// 空スライスに直接append
+		dst := append([]int{}, items...)
+		voidfn(dst)
+	}
+}
+
+// 方法3: 容量をリセットしてからappendする方法
+func BenchmarkCapacityResetAppend(b *testing.B) {
+	var (
+		items = make([]int, 0, NumItems)
+	)
+	for i := range NumItems {
+		items = append(items, i)
+	}
+
+	for b.Loop() {
+		var (
+			dst = make([]int, 0, NumItems)
+		)
+		// スライスの長さと容量を0にリセットしてからappend
+		dst = append(dst[:0:0], items...)
+		voidfn(dst)
+	}
+}
+
+// 方法4: 従来のcopy()を使った方法
+func BenchmarkBuiltinCopy(b *testing.B) {
+	var (
+		items = make([]int, 0, NumItems)
+	)
+	for i := range NumItems {
+		items = append(items, i)
+	}
+
+	for b.Loop() {
+		// 正しい長さでスライスを作成
+		dst := make([]int, len(items))
+		// copyを使用してコピー（最も効率的）
+		copy(dst, items)
+		voidfn(dst)
+	}
+}
+
+// 方法5: slices.Clone()を使った方法
+func BenchmarkSlicesClone(b *testing.B) {
+	var (
+		items = make([]int, 0, NumItems)
+	)
+	for i := range NumItems {
+		items = append(items, i)
+	}
+
+	for b.Loop() {
+		// slices.Cloneを使用（内部でmake + copyを実行）
+		dst := slices.Clone(items)
+		voidfn(dst)
+	}
+}
+
+/*
+$ task
+task: [default] go test -bench . -benchmem -count 3 | tee bench.txt
+goos: linux
+goarch: amd64
+pkg: github.com/devlights/try-golang/examples/benchmarks/03.slice_clone
+cpu: Intel(R) Core(TM) Ultra 5 125H
+BenchmarkPreAllocatedAppend-18               134           8116946 ns/op        41943090 B/op          1 allocs/op
+BenchmarkPreAllocatedAppend-18               150           7897353 ns/op        41943049 B/op          1 allocs/op
+BenchmarkPreAllocatedAppend-18               153           7697424 ns/op        41943048 B/op          1 allocs/op
+BenchmarkEmptySliceAppend-18                 234           5234496 ns/op        41943072 B/op          1 allocs/op
+BenchmarkEmptySliceAppend-18                 246           4907417 ns/op        41943081 B/op          1 allocs/op
+BenchmarkEmptySliceAppend-18                 174           6013573 ns/op        41943054 B/op          1 allocs/op
+BenchmarkCapacityResetAppend-18              153           7743873 ns/op        83886086 B/op          2 allocs/op
+BenchmarkCapacityResetAppend-18              152           7804002 ns/op        83886092 B/op          2 allocs/op
+BenchmarkCapacityResetAppend-18              150           7891077 ns/op        83886088 B/op          2 allocs/op
+BenchmarkBuiltinCopy-18                      151           7796522 ns/op        41943046 B/op          1 allocs/op
+BenchmarkBuiltinCopy-18                      152           7813188 ns/op        41943050 B/op          1 allocs/op
+BenchmarkBuiltinCopy-18                      158           7525531 ns/op        41943050 B/op          1 allocs/op
+BenchmarkSlicesClone-18                      255           5556701 ns/op        41943047 B/op          1 allocs/op
+BenchmarkSlicesClone-18                      187           6222304 ns/op        41943047 B/op          1 allocs/op
+BenchmarkSlicesClone-18                      204           5850473 ns/op        41943044 B/op          1 allocs/op
+PASS
+ok      github.com/devlights/try-golang/examples/benchmarks/03.slice_clone      17.955s
+task: [default] benchstat bench.txt
+goos: linux
+goarch: amd64
+pkg: github.com/devlights/try-golang/examples/benchmarks/03.slice_clone
+cpu: Intel(R) Core(TM) Ultra 5 125H
+                       │  bench.txt   │
+                       │    sec/op    │
+PreAllocatedAppend-18    7.897m ± ∞ ¹
+EmptySliceAppend-18      5.234m ± ∞ ¹
+CapacityResetAppend-18   7.804m ± ∞ ¹
+BuiltinCopy-18           7.797m ± ∞ ¹
+SlicesClone-18           5.850m ± ∞ ¹
+geomean                  6.816m
+¹ need >= 6 samples for confidence interval at level 0.95
+
+                       │   bench.txt   │
+                       │     B/op      │
+PreAllocatedAppend-18    40.00Mi ± ∞ ¹
+EmptySliceAppend-18      40.00Mi ± ∞ ¹
+CapacityResetAppend-18   80.00Mi ± ∞ ¹
+BuiltinCopy-18           40.00Mi ± ∞ ¹
+SlicesClone-18           40.00Mi ± ∞ ¹
+geomean                  45.95Mi
+¹ need >= 6 samples for confidence interval at level 0.95
+
+                       │  bench.txt  │
+                       │  allocs/op  │
+PreAllocatedAppend-18    1.000 ± ∞ ¹
+EmptySliceAppend-18      1.000 ± ∞ ¹
+CapacityResetAppend-18   2.000 ± ∞ ¹
+BuiltinCopy-18           1.000 ± ∞ ¹
+SlicesClone-18           1.000 ± ∞ ¹
+geomean                  1.149
+¹ need >= 6 samples for confidence interval at level 0.95
+task: [default] rm -f bench.txt
+*/
+
+/*
+	結果として、何度試しても空スライスにAppendする方法が最も速かった。
+	つまり、
+		dst := append([]int{}, items...)
+	が速い。
+	slices.Clone()は、内部で上記を行っている
+		// Avoid s[:0:0] as it leads to unwanted liveness when cloning a
+		// zero-length slice of a large array; see https://go.dev/issue/68488.
+		return append(S{}, s...)
+*/