OneOfOne / lfchan

A scalable lock-free channel.

lfchan

--

A scalable lock-free channel.

• Supports graceful closing.
• Supports blocking and non-blocking operations.
• Supports select.
• Scales with the number of cores.

Install

go get github.com/OneOfOne/lfchan

Usage

import (
	"fmt"

	"github.com/OneOfOne/lfchan"
)

func main() {
	ch := lfchan.New() // or
	// ch := lfchan.NewSize(10) // buffered channel
	go ch.Send("hello", true)
	fmt.Printf("%s world", ch.Recv(true).(string))
}

Generate a typed channel

Generate the package:

go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" type [pkgName or . to embed the chan in the current package]

# primitve type
go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" string internal/stringChan

# or for using a non-native type
go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" github.com/OneOfOne/cmap.CMap internal/cmapChan

go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" github.com/OneOfOne/cmap.CMap

Use it in your code:

typed sub package

package main

// go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" string internal/stringChan

import (
	"fmt"

	"github.com/YOU/internal/stringChan"
)

func main() {
	ch := stringChan.New() // or
	// ch := stringChan.NewSize(10) // buffered channel
	go func() {
		go ch.Send("lfchan", true)
		ch.Send("hello", true)
	}()
	for s, ok := ch.Recv(true); ok; s, ok = ch.Recv(true) {
		fmt.Print(s, " ")
	}
	fmt.Println()
}

embed the type directly

package main

// go run "$GOPATH/src/github.com/OneOfOne/lfchan/gen.go" "[]*node" .

import (
	"fmt"
)

type node struct {
	v int
}

func main() {
	// notice how for embeded types the new func is called "new[Size]{TypeName}Chan()
	ch := newNodeChan() // or
	// ch := newSizeNodeChan(10) // buffered channel
	go func() {
		for i := 0; i < 10; i++ {
			ch.Send([]*Node{{i}, {i*i}}, true)
		}
	}()
	for ns, ok := ch.Recv(true); ok; ns, ok = ch.Recv(true) {
		for i, n := range ns {
			fmt.Println(i, n.v)
		}
	}
}

Known issues

• Doesn't scale correctly on true SMP systems (issue #3).

• Under high concurrency, ch.Len() can return -1 (issue #2) Fixed by commit bdddd90.

• typed channels can't handle zero value primitve types correctly, for example it can't handle sending 0 on an int channel Fixed.

• gen.go can't handle maps to non-native types.

Benchmark

# Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz
# Linux 4.4.5 x86_64

➜ go test -bench=. -benchmem -cpu 1,4,8 -benchtime 10s
BenchmarkLFChan         100000000              157 ns/op               8 B/op          1 allocs/op
BenchmarkLFChan-4       100000000              187 ns/op               8 B/op          1 allocs/op
BenchmarkLFChan-8       100000000              168 ns/op               8 B/op          1 allocs/op

BenchmarkChan           200000000             96.2 ns/op               8 B/op          1 allocs/op
BenchmarkChan-4         50000000               244 ns/op               8 B/op          1 allocs/op
BenchmarkChan-8         50000000               323 ns/op               8 B/op          1 allocs/op

PASS
ok      github.com/OneOfOne/lfchan      124.067s

check issue #3 for more benchmarks and updates.

FAQ

Why are you using runtime.Gosched?

• Sadly, it is the only clean way to release the scheduler in a tight loop, Go doesn't provide any other way to yield, time.Sleep causes random allocations at times. sync/atomic.Value has access to internal funcs which can control the scheduler, however user code can't do that.

Isn't using a spinlock bad for the CPU?

• Yes and no, using the spinlock and few sleeps in the code makes it very efficient even under idle conditions.

License

This project is released under the Apache v2. licence. See LICENCE for more details.