这是本系列文章的第二篇,第一篇在此golang并发三板斧系列之一:channel用于通信和同步。
前文描述了手工作坊的时代,即老师傅带着小学徒并发地做一项工作,现在我们准备进入工业时代。
Pipeline模型
Pipeline即流水线模型,这在现代工业是很常见的。模型分为数个阶段,每个阶段干不同的事情,但可以并行地去做。以造拖拉机为例来解释流水线的工作方式,假设装配一辆汽车需要四个步骤:
- 第一步冲压:制作车身外壳和底盘等部件。
- 第二步焊接:将冲压成形后的各部件焊接成车身。
- 第三步涂装:将车身等主要部件清洗、化学处理、打磨、喷漆和烘干。
- 第四步总装:将各部件(包括发动机和向外采购的零部件)组装成车。
比如要造一百辆拖拉机,如果每个阶段都等前一阶段的一百辆完成才开工,是对生产线的极大浪费。现代工业的流水线做法是每个阶段同时开工,在时间上并行起来。流水线的概念在计算机世界中也很普遍,拥有流水线的CPU可以在一个时钟周期内完成一条指令,而不是等待取指令、译码、取操作数、执行四个阶段才能完成一条指令:
golang的设计者们吸收了这一经典概念,使用channel把前后数个阶段串联起来,形成一个流水线:
func gen(nums ...int) <-chan int {
out := make(chan int)
go func() {
for _, n := range nums {
out <- n
}
close(out)
}()
return out
}
func sq(in <-chan int) <-chan int {
out := make(chan int)
go func() {
for n := range in {
out <- n * n
}
close(out)
}()
return out
}
func TestPipeline(t *testing.T) {
c := gen(2, 3, 4)
out := sq(c)
for n := range out {
log.Print(n)
}
}
由于sq
函数的入参和出参一样,故可以增加无数个阶段写为:
func TestPipelines(t *testing.T) {
for n := range sq(sq(sq(gen(1, 2, 3, 4)))) {
log.Print(n)
}
}
FAN模型
可以说FAN模型是流水线的一种改进。可以观察到上述的流水线模型,每个阶段只起了一个goroutine,但是现实造拖拉机的时候,在组装轮子的时候可以4个工人一起上,这又是提高了并发。golang吸取了这种模型,在任务分发阶段,多个goroutine从同一个channel读取数据,直到关闭,称为FAN-OUT模型;在结果收集阶段,单个goroutine从多个channel读取数据,直到关闭,称为FAN-IN模型:
[图片上传失败...(image-d2aaa-1558316671133)]
func merge(cs ...<-chan int) <-chan int {
var wg sync.WaitGroup
out := make(chan int)
output := func(c <-chan int) {
for n := range c {
out <- n
}
wg.Done()
}
wg.Add(len(cs))
for _, c := range cs {
go output(c)
}
go func() {
wg.Wait()
close(out)
}()
return out
}
func TestFan(t *testing.T) {
c := gen(2, 3, 4)
out1 := sq(c)
out2 := sq(c)
for n := range merge(out1, out2) {
log.Print(n)
}
}
并行起来之后,最终结果的顺序是不可控的:
2019/02/22 21:39:27 9
2019/02/22 21:39:27 16
2019/02/22 21:39:27 4
Process finished with exit code 0
以上模型能提升并行效率吗
CPU密集型
用浮点数的幂计算模拟CPU-BOUND,设计了如下模型用于比较:
//功能函数
func benchmarkList() []float64 {
list := make([]float64, MAX)
for n := 0; n < MAX; n++ {
list[n] = float64(n)
}
return list
}
func cpubound(n float64) float64 {
return math.Pow(3.1415926, n)
}
func genChan(nums []float64) <-chan float64 {
out := make(chan float64)
go func() {
for _, n := range nums {
out <- n
}
close(out)
}()
return out
}
func genChanBuffer(nums []float64) <-chan float64 {
out := make(chan float64, BUFFERSIZE)
go func() {
for _, n := range nums {
out <- n
}
close(out)
}()
return out
}
func cpuChan(in <-chan float64) <-chan float64 {
out := make(chan float64)
go func() {
for n := range in {
out <- cpubound(n)
}
close(out)
}()
return out
}
func cpuChanBuffer(in <-chan float64) <-chan float64 {
out := make(chan float64, BUFFERSIZE)
go func() {
for n := range in {
out <- cpubound(n)
}
close(out)
}()
return out
}
func mergeChan(cs ...<-chan float64) <-chan float64 {
var wg sync.WaitGroup
out := make(chan float64)
output := func(c <-chan float64) {
for n := range c {
out <- n
}
wg.Done()
}
wg.Add(len(cs))
for _, c := range cs {
go output(c)
}
go func() {
wg.Wait()
close(out)
}()
return out
}
func mergeChanBuffer(cs ...<-chan float64) <-chan float64 {
var wg sync.WaitGroup
out := make(chan float64, BUFFERSIZE)
output := func(c <-chan float64) {
for n := range c {
out <- n
}
wg.Done()
}
wg.Add(len(cs))
for _, c := range cs {
go output(c)
}
go func() {
wg.Wait()
close(out)
}()
return out
}
//测试函数
func BenchmarkCPUSequential(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
var sum float64
for _, n := range list {
sum += cpubound(n)
}
}
}
func BenchmarkCPUPipeline(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChan(list)
out := cpuChan(c)
var sum float64
for n := range out {
sum += n
}
}
}
func BenchmarkCPUPipelineBuffer(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChanBuffer(list)
out := cpuChanBuffer(c)
var sum float64
for n := range out {
sum += n
}
}
}
func BenchmarkCPUFan(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChan(list)
gonum := runtime.NumCPU() / 2
outs := make([]<-chan float64, gonum)
for i := 0; i < gonum; i++ {
outs[i] = cpuChan(c)
}
var sum float64
for n := range mergeChan(outs...) {
sum += n
}
}
}
func BenchmarkCPUFanBuffer(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChanBuffer(list)
gonum := runtime.NumCPU() / 2
outs := make([]<-chan float64, gonum)
for i := 0; i < gonum; i++ {
outs[i] = cpuChanBuffer(c)
}
var sum float64
for n := range mergeChanBuffer(outs...) {
sum += n
}
}
}
func BenchmarkCPUParallelize(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
gonum := runtime.NumCPU()
var sum float64
num := len(list)
stride := num / gonum
var wg sync.WaitGroup
wg.Add(gonum)
var mux sync.Mutex
for g := 0; g < gonum; g++ {
go func(g int) {
start := g * stride
end := start + stride
if g == gonum-1 {
end = num
}
var sumin float64
for _, n := range list[start:end] {
sumin += cpubound(n)
}
mux.Lock()
sum += sumin
mux.Unlock()
wg.Done()
}(g)
}
wg.Wait()
}
}
设MAX=1000000,BUFFERSIZE=1000。结果让人大跌眼镜,无论是Pipeline模型还是Fan模型,都比不上普通的串行,只有普通的并发模型能有效提升:
[baixiao@localhost go_concurrency]$ GOGC=off go test -cpu 1,8 -run none -bench CPU -benchtime 3s
goos: linux
goarch: amd64
BenchmarkCPUSequential 50 95148037 ns/op
BenchmarkCPUSequential-8 30 101450384 ns/op
BenchmarkCPUPipeline 10 512093124 ns/op
BenchmarkCPUPipeline-8 5 864946495 ns/op
BenchmarkCPUPipelineBuffer 20 219850707 ns/op
BenchmarkCPUPipelineBuffer-8 10 370302165 ns/op
BenchmarkCPUFan 5 715223945 ns/op
BenchmarkCPUFan-8 5 913448396 ns/op
BenchmarkCPUFanBuffer 10 320494600 ns/op
BenchmarkCPUFanBuffer-8 10 427863250 ns/op
BenchmarkCPUParallelize 100 95482003 ns/op
BenchmarkCPUParallelize-8 200 19398520 ns/op
PASS
ok _/home/baixiao/go_concurrency 77.085s
可以得出以下结论:
- Sequential完爆Pipeline和Fan
- Pipeline和Fan在多核下均弱于单核,因为系统瓶颈根本不在并行上,而是channel造成的阻塞
- 给channel加了buffer之后,PipelineBuffer优于Pipeline、FanBuffer优于Fan,因为channel的阻塞减弱了
- 多核下的Parallelize:在座的各位都是垃圾
IO密集型
用随机sleep模拟IO-BOUND,设计了如下模型用于比较:
//功能函数
func iobound(n float64) float64 {
time.Sleep(time.Duration(rand.Intn(10)) * time.Millisecond)
return n
}
func ioChan(in <-chan float64) <-chan float64 {
out := make(chan float64)
go func() {
for n := range in {
out <- iobound(n)
}
close(out)
}()
return out
}
func ioChanBuffer(in <-chan float64) <-chan float64 {
out := make(chan float64, BUFFERSIZE)
go func() {
for n := range in {
out <- iobound(n)
}
close(out)
}()
return out
}
//测试函数
func BenchmarkIOSequential(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
var sum float64
for _, n := range list {
sum += iobound(n)
}
}
}
func BenchmarkIOPipeline(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChan(list)
out := ioChan(c)
var sum float64
for n := range out {
sum += n
}
}
}
func BenchmarkIOPipelineBuffer(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChanBuffer(list)
out := ioChanBuffer(c)
var sum float64
for n := range out {
sum += n
}
}
}
func BenchmarkIOFan(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChan(list)
gonum := runtime.NumCPU() / 2
outs := make([]<-chan float64, gonum)
for i := 0; i < gonum; i++ {
outs[i] = ioChan(c)
}
var sum float64
for n := range mergeChan(outs...) {
sum += n
}
}
}
func BenchmarkIOFanBuffer(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
c := genChanBuffer(list)
gonum := runtime.NumCPU() / 2
outs := make([]<-chan float64, gonum)
for i := 0; i < gonum; i++ {
outs[i] = ioChanBuffer(c)
}
var sum float64
for n := range mergeChanBuffer(outs...) {
sum += n
}
}
}
func BenchmarkIOParallelize(b *testing.B) {
for i := 0; i < b.N; i++ {
list := benchmarkList()
gonum := runtime.NumCPU()
var sum float64
num := len(list)
stride := num / gonum
var wg sync.WaitGroup
wg.Add(gonum)
var mux sync.Mutex
for g := 0; g < gonum; g++ {
go func(g int) {
start := g * stride
end := start + stride
if g == gonum-1 {
end = num
}
var sumin float64
for _, n := range list[start:end] {
sumin += iobound(n)
}
mux.Lock()
sum += sumin
mux.Unlock()
wg.Done()
}(g)
}
wg.Wait()
}
}
设MAX=100,BUFFERSIZE=1000。
[baixiao@localhost go_concurrency]$ GOGC=off go test -cpu 1,8 -run none -bench IO -benchtime 3s
goos: linux
goarch: amd64
BenchmarkIOSequential 10 441609349 ns/op
BenchmarkIOSequential-8 10 457200927 ns/op
BenchmarkIOPipeline 10 454147034 ns/op
BenchmarkIOPipeline-8 10 467264740 ns/op
BenchmarkIOPipelineBuffer 10 456459492 ns/op
BenchmarkIOPipelineBuffer-8 10 452286832 ns/op
BenchmarkIOFan 100 36995490 ns/op
BenchmarkIOFan-8 100 36950275 ns/op
BenchmarkIOFanBuffer 100 37555702 ns/op
BenchmarkIOFanBuffer-8 100 36851459 ns/op
BenchmarkIOParallelize 50 88653231 ns/op
BenchmarkIOParallelize-8 50 87061568 ns/op
PASS
ok _/home/baixiao/go_concurrency 54.007s
可以得出以下结论:
- 在IO密集的情况下,Fan模型吊打所有
- channel带buffer没有效率提升
- 所有的模型,在多核下都没有提升
结论是?
- 不带buffer的channel由于「强同步」特性,无法提高并行,甚至拖累效率
- CPU密集型的场景,多核并行能提升效率
- IO密集型的场景,多核并行不能提升效率
- Pipeling模型有何用途?我没看出来
- Waitgroup模型在CPU密集型场景有优势
- Fan模型在IO密集型场景有优势
坑:runtime.NumCPU()不会随着runtime.GOMAXPROCS()改变,前者代表的是系统全部的核数,后者代表的是可同时使用的核数
goroutine池
pool模型
设计一个pool,需要考虑几个方面:输入是什么,做什么事情,多少worker一起执行?
现抽象出一个goroutine pool的模型代码,可以自定义输入类型,执行函数,worker数量。
func genPoolChanBuffer(nums []float64) <-chan interface{} {
out := make(chan interface{}, BUFFERSIZE)
go func() {
for _, n := range nums {
out <- n
}
close(out)
}()
return out
}
type Handler func(*Work)
type Work struct {
input interface{}
}
type Pool struct {
channum int
workernum int
wg *sync.WaitGroup
ch chan Work
Func Handler
}
func (p Pool) RunWorker() {
for i := 0; i < p.workernum; i++ {
p.wg.Add(1)
go func() {
defer p.wg.Done()
for work := range p.ch {
p.Func(&work)
}
}()
}
}
func (p Pool) FeedWorker(in <-chan interface{}) {
go func() {
for n := range in {
work := Work{
input: n,
}
p.ch <- work
}
close(p.ch)
}()
}
func (p Pool) Wait() {
p.wg.Wait()
}
func InitPool(channum, workernum int, f Handler) *Pool {
return &Pool{
channum: channum,
workernum: workernum,
wg: &sync.WaitGroup{},
ch: make(chan Work, channum),
Func: f,
}
}
对比测试
设计了四个对比测试,观察在CPU密集型和IO密集型的情况下该pool的表现,另外还旨在探索什么情况下worker的数量会越多越好?带Min的测试中设置gonum
为系统核数的一半,带Max的测试中设置gonum
为系统核数的十倍,gonum
即为worker数量。
//测试函数
func BenchmarkCPUPool(b *testing.B) {
channum := 100
gonum := runtime.NumCPU()
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
cpubound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
func BenchmarkCPUPoolMin(b *testing.B) {
channum := 100
gonum := runtime.NumCPU() / 2
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
cpubound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
func BenchmarkCPUPoolMax(b *testing.B) {
channum := 100
gonum := 10 * runtime.NumCPU()
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
cpubound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
func BenchmarkIOPool(b *testing.B) {
channum := 100
gonum := runtime.NumCPU()
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
iobound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
func BenchmarkIOPoolMin(b *testing.B) {
channum := 100
gonum := runtime.NumCPU() / 2
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
iobound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
func BenchmarkIOPoolMax(b *testing.B) {
channum := 100
gonum := 10 * runtime.NumCPU()
for i := 0; i < b.N; i++ {
f := func(w *Work) {
if v, ok := w.input.(float64); ok {
iobound(v)
}
}
list := benchmarkList()
c := genPoolChanBuffer(list)
p := InitPool(channum, gonum, f)
p.RunWorker()
p.FeedWorker(c)
p.Wait()
}
}
设MAX=100,BUFFERSIZE=1000。
[baixiao@localhost go_concurrency]$ GOGC=off go test -cpu 1,8 -run none -bench Pool -benchtime 3s
goos: linux
goarch: amd64
BenchmarkCPUPool 100000 38070 ns/op
BenchmarkCPUPool-8 50000 77872 ns/op
BenchmarkCPUPoolMin 100000 33286 ns/op
BenchmarkCPUPoolMin-8 50000 71690 ns/op
BenchmarkCPUPoolMax 30000 143250 ns/op
BenchmarkCPUPoolMax-8 20000 281619 ns/op
BenchmarkIOPool 200 21382667 ns/op
BenchmarkIOPool-8 200 21467617 ns/op
BenchmarkIOPoolMin 100 37068480 ns/op
BenchmarkIOPoolMin-8 100 37350682 ns/op
BenchmarkIOPoolMax 500 9438800 ns/op
BenchmarkIOPoolMax-8 500 9519574 ns/op
PASS
ok _/home/baixiao/go_concurrency 64.149s
可以得出以下结论:
- CPU密集型场景中多核下均弱于单核
- CPU密集型场景中worker数量太多只能起反作用
- IO密集型场景中多核并行不能提升效率
- IO密集型场景中worker数量在一定范围内能有效提升效率
- Pool模型由于用到了channel,多核都不能提升效率
channel是个好东西?
在第一篇里,我们讲到channel是goroutine之间通信和同步的重要工具,也是golang中重要的关键字之一,说明golang的设计者们很看重这个特性。
但是实际上channel的性能较一般,分析源码可知,channel中的数据无论读写都会加mutex锁,造成高并发时的较大瓶颈,这个从我们的对比测试中也都可以看出来。
上图来自文章。
另外,channel目前都还有整个社群都无法调优的问题,比如runtime: select on a shared channel is slow with many Ps。
本篇的模型可以类比为你现在是个企业主,开工厂进行工业生产了。
所有代码都在https://github.com/baixiaoustc/go_concurrency/blob/master/second_post_test.go中能找到。