stack-unwind-overhead

2025-07-12 23:50:13 +02:00 · 2021-04-04 14:39:53 +02:00
parent f8f6e88e3a
commit b393c68281
4 changed files with 695 additions and 0 deletions
--- a/examples/stack-unwind-overhead/README.md
+++ b/examples/stack-unwind-overhead/README.md
@ -0,0 +1,46 @@
+# stack-unwind-overhead
+
+This directory contains benchmarks to explore which factors impact stack unwinding in Go. It's informed by an analysis of the `gopclntab` unwinding implementation.
+
+## Result 1: O(N) Stack Depth
+
+The benchmark below shows that stack unwinding has O(N) complexity with regard to the number of call frames on the stack:
+
+```
+BenchmarkStackDepth/8-12    	2208600	      547.2 ns/op
+BenchmarkStackDepth/16-12   	1447922	      810.8 ns/op
+BenchmarkStackDepth/32-12   	 915291	     1338 ns/op
+BenchmarkStackDepth/64-12   	 488719	     2366 ns/op
+BenchmarkStackDepth/128-12  	 264735	     4462 ns/op
+BenchmarkStackDepth/256-12  	 137575	     8643 ns/op
+BenchmarkStackDepth/512-12  	  68355	    17316 ns/op
+BenchmarkStackDepth/1024-12 	  34710	    34810 ns/op
+```
+
+## Result 2: O(N) Function Size
+
+Perhaps suprisingly, stack unwinding is also O(N) with regard to the size of the generated machine code for the function:
+
+```
+BenchmarkFunctionSize/1-12  	2562176	      462.8 ns/op
+BenchmarkFunctionSize/2-12  	2509465	      484.7 ns/op
+BenchmarkFunctionSize/4-12  	2356609	      504.6 ns/op
+BenchmarkFunctionSize/8-12  	2095870	      568.3 ns/op
+BenchmarkFunctionSize/16-12 	1778889	      669.7 ns/op
+BenchmarkFunctionSize/32-12 	1396009	      856.0 ns/op
+BenchmarkFunctionSize/64-12 	 943807	     1269 ns/op
+BenchmarkFunctionSize/128-12         	 516487	     2271 ns/op
+BenchmarkFunctionSize/256-12         	 277821	     4490 ns/op
+```
+
+## Disclaimer
+
+YMMV, and especially the function size also depends on the program counter at which the function is being unwound. All tests were done on my local machine:
+
+```
+go test -bench .
+goos: darwin
+goarch: amd64
+pkg: github.com/felixge/go-profiler-notes/examples/stack-unwind-overhead
+cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
+```
--- a/examples/stack-unwind-overhead/go.mod
+++ b/examples/stack-unwind-overhead/go.mod
@ -0,0 +1,3 @@
+module github.com/felixge/go-profiler-notes/examples/stack-unwind-overhead
+
+go 1.16
--- a/examples/stack-unwind-overhead/main.go
+++ b/examples/stack-unwind-overhead/main.go
@ -0,0 +1,18 @@
+// Package main is a code generator that creates functions of different size.
+package main
+
+import (
+	"fmt"
+	"os"
+)
+
+func main() {
+	if err := run(); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+}
+
+func run() error {
+	return nil
+}
--- a/examples/stack-unwind-overhead/main_test.go
+++ b/examples/stack-unwind-overhead/main_test.go
@ -0,0 +1,628 @@
+package main
+
+import (
+	"fmt"
+	"runtime"
+	"testing"
+)
+
+// BenchmarkStackDepth measures the impact of stack depth on the time it takes
+// to create a stack trace.
+func BenchmarkStackDepth(b *testing.B) {
+	for d := 8; d <= 1024; d = d * 2 {
+		b.Run(fmt.Sprintf("%d", d), func(b *testing.B) {
+			callers := make([]uintptr, d)
+			atStackDepth(d, func() {
+				for i := 0; i < b.N; i++ {
+					if n := runtime.Callers(0, callers); n != d {
+						b.Fatalf("got=%d want=%d", n, d)
+					}
+				}
+			})
+		})
+	}
+}
+
+// atStackDepth recursively calls itself until the number of stack frames
+// reaches the desired depth, and then calls fn.
+func atStackDepth(depth int, fn func()) {
+	pcs := make([]uintptr, depth+10)
+	n := runtime.Callers(1, pcs)
+	if n > depth {
+		panic("depth exceeded")
+	} else if n < depth {
+		atStackDepth(depth, fn)
+		return
+	}
+
+	fn()
+}
+
+func BenchmarkFunctionSize(b *testing.B) {
+	m := map[int]func(func()){
+		1:   dostuff1,
+		2:   dostuff2,
+		4:   dostuff4,
+		8:   dostuff8,
+		16:  dostuff16,
+		32:  dostuff32,
+		64:  dostuff64,
+		128: dostuff128,
+		256: dostuff256,
+	}
+
+	var callers = make([]uintptr, 32)
+	for s := 1; s <= 256; s = s * 2 {
+		b.Run(fmt.Sprintf("%d", s), func(b *testing.B) {
+			m[s](func() {
+				for i := 0; i < b.N; i++ {
+					runtime.Callers(0, callers)
+				}
+			})
+		})
+	}
+
+	//b.Run("long", func(b *testing.B) {
+	//long(func() {
+	//for i := 0; i < b.N; i++ {
+	//runtime.Callers(0, callers)
+	//}
+	//})
+	//})
+}
+
+// dostuff is supposed to generate a bunch of machine code that will hopefully
+// be inlined into its callers.
+func dostuff() int {
+	m := map[int]string{}
+	m[0] = "foo"
+	m[100] = "bar"
+	return len(m)
+}
+
+func dostuff1(fn func()) {
+	dostuff()
+	fn()
+}
+
+func dostuff2(fn func()) {
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff4(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff8(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff16(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff32(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff64(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff128(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}
+
+func dostuff256(fn func()) {
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	dostuff()
+	fn()
+}