cp benchmarks

cp.go

package main

import (
	"flag"
	"fmt"
	"io"
	"os"
	"path"
	"runtime"
)

var offchan chan int64
var zerochan chan int

func copy(from, to string, todir bool) bool {
	if todir {
		_, file := path.Split(from)
		to = to + "/" + file
	}

	dirb, err := os.Stat(from)
	if err != nil {
		fmt.Printf("can't stat %s: %v\n", from, err)
		return true
	}
	tob, err := os.Stat(to)
	if err == nil {
		if sameFile(dirb.Sys(), tob.Sys()) {
			fmt.Printf("%s and %s are the same file\n", from, to)
			return true
		}
	}

	if dirb.IsDir() {
		fmt.Printf("%s is a directory\n", from)
		return true
	}

	mode := dirb.Mode() & 0777
	f, err := os.Open(from)
	if err != nil {
		fmt.Printf("can't open %s: %v\n", from, err)
		return true
	}
	defer f.Close()

	t, err := os.OpenFile(to, os.O_RDWR|os.O_CREATE|os.O_TRUNC, mode)
	if err != nil {
		fmt.Printf("can't create %s: %v\n", to, err)
		f.Close()
		return true
	}
	defer t.Close()
	return copy1(f, t, from, to)
}

func copy1(f, t *os.File, from, to string) (ret bool) {
	zerochan <- 0
	fail := make(chan bool, Nwork)

	for i := 0; i < Nwork; i++ {
		go worker(f, t, from, to, fail)
	}
	for i := 0; i < Nwork; i++ {
		end := <-fail
		if end == true {
			ret = true
		}
	}
	return
}

func worker(f, t *os.File, from, to string, fail chan bool) {
	buf := make([]byte, BufSize)
	var bp []byte

	l := len(buf)
	bp = buf[0:]
	o := <-offchan
	for {
		n, err := f.ReadAt(bp, o)
		if err != nil && err != io.EOF {
			fmt.Printf("reading %s at %v: %v\n", from, o, err)
			fail <- true
			return
		}
		if n == 0 {
			break
		}

		nb := bp[0:n]
		n, err = t.WriteAt(nb, o)
		if err != nil {
			fmt.Printf("writing %s: %v\n", to, err)
			fail <- true
			return
		}
		bp = buf[n:]
		o += int64(n)
		l -= n
		if l == 0 {
			l = len(buf)
			bp = buf[0:]
			o = <-offchan
		}
	}
	fail <- false
}

func nextoff() {
	off := int64(0)
	for {
		select {
		case <-zerochan:
			off = 0
		case offchan <- off:
			off += BufSize
		}
	}
}

func usage() {
	fmt.Printf("usage: cp [-w workers] from to\n")
	os.Exit(1) // sysfatal
}

var Nwork int
var BufSize int64
var Nthreads int

func main() {
	todir := false

	flag.IntVar(&Nwork, "w", 16, "number of worker goroutines")
	flag.Int64Var(&BufSize, "bs", 4, "buffer size in KB")
	flag.IntVar(&Nthreads, "t", 1, "number of OS threads (GOMAXPROCS). 0=number of CPU cores")
	flag.Parse()
	BufSize = BufSize << 10
	switch flag.NArg() {
	case 0, 1:
		usage()
	}

	if Nthreads == 0 {
		Nthreads = runtime.NumCPU()
	}
	runtime.GOMAXPROCS(Nthreads)

	files := flag.Args()
	lf := files[len(files)-1]
	lfdir, err := os.Stat(lf)
	if err == nil {
		todir = lfdir.IsDir()
	}
	if flag.NArg() > 2 && todir == false {
		fmt.Printf("not a directory: %s\n", lf)
		os.Exit(1) // sysfatal
	}

	offchan = make(chan int64, 0)
	zerochan = make(chan int, 0)
	go nextoff()

	failed := false
	for i := 0; i < flag.NArg()-1; i++ {
		if copy(files[i], lf, todir) {
			failed = true
		}
	}
	if failed {
		os.Exit(2)
	}
	return
}

gistfile1.txt

# Story: f2f on irc.freenode#go-nuts posted his program
# http://mirtchovski.com/go/cp/
# Which implements plan9 cp using concurrent pread/pwrite blocks.

$ cat drop-caches 
#!/bin/bash -e

sync
sleep 1s
sync
echo 3 >/proc/sys/vm/drop_caches


$ for i in {1..5}; do sudo ./drop-caches ; time ./f2f-cp -w=1 -t=1 -bs=8 200m 200m.1 ; cmp 200m 200m.1 || echo error; rm -f 200m.1 ; done
0.30user 1.31system 0:04.89elapsed 33%CPU (0avgtext+0avgdata 1200maxresident)k
411912inputs+409600outputs (9major+345minor)pagefaults 0swaps
0.26user 1.37system 0:04.98elapsed 32%CPU (0avgtext+0avgdata 1196maxresident)k
411912inputs+409600outputs (9major+343minor)pagefaults 0swaps
0.32user 1.30system 0:04.98elapsed 32%CPU (0avgtext+0avgdata 1192maxresident)k
411912inputs+409600outputs (9major+344minor)pagefaults 0swaps
0.26user 1.30system 0:04.89elapsed 32%CPU (0avgtext+0avgdata 1200maxresident)k
411912inputs+409600outputs (9major+346minor)pagefaults 0swaps
0.25user 1.25system 0:05.02elapsed 30%CPU (0avgtext+0avgdata 1196maxresident)k
411912inputs+409600outputs (9major+343minor)pagefaults 0swaps

$ for i in {1..5}; do sudo ./drop-caches ; time ./f2f-cp -w=4 -t=4 -bs=8 200m 200m.1 ; cmp 200m 200m.1 || echo error; rm -f 200m.1 ; done
0.09user 0.81system 0:05.05elapsed 18%CPU (0avgtext+0avgdata 1320maxresident)k
411912inputs+409600outputs (9major+378minor)pagefaults 0swaps
0.11user 0.79system 0:04.95elapsed 18%CPU (0avgtext+0avgdata 1300maxresident)k
411912inputs+409600outputs (9major+373minor)pagefaults 0swaps
0.10user 0.80system 0:05.02elapsed 18%CPU (0avgtext+0avgdata 1304maxresident)k
411912inputs+409600outputs (9major+374minor)pagefaults 0swaps
0.11user 0.79system 0:05.08elapsed 17%CPU (0avgtext+0avgdata 1316maxresident)k
411912inputs+409600outputs (9major+378minor)pagefaults 0swaps
0.13user 0.77system 0:05.27elapsed 17%CPU (0avgtext+0avgdata 1316maxresident)k
411912inputs+409600outputs (9major+376minor)pagefaults 0swaps

$ for i in {1..5}; do sudo ./drop-caches ; time ./f2f-cp -w=4 -t=1 -bs=8 200m 200m.1 ; cmp 200m 200m.1 || echo error; rm -f 200m.1 ; done
0.16user 0.89system 0:05.30elapsed 20%CPU (0avgtext+0avgdata 1324maxresident)k
411912inputs+409600outputs (9major+393minor)pagefaults 0swaps
0.17user 0.88system 0:05.25elapsed 20%CPU (0avgtext+0avgdata 1320maxresident)k
411912inputs+409600outputs (9major+392minor)pagefaults 0swaps
0.15user 0.91system 0:05.06elapsed 21%CPU (0avgtext+0avgdata 1324maxresident)k
411912inputs+409600outputs (9major+393minor)pagefaults 0swaps
0.15user 0.91system 0:05.11elapsed 20%CPU (0avgtext+0avgdata 1328maxresident)k
411912inputs+409600outputs (9major+394minor)pagefaults 0swaps
0.16user 0.90system 0:05.40elapsed 19%CPU (0avgtext+0avgdata 1332maxresident)k
411912inputs+409600outputs (9major+395minor)pagefaults 0swaps