假设服务器端提供无限的服务能力，客户端如何达到最高的QPS？如果使用多线程来提高并发，线程数量会成为瓶颈。是否可以通过协程来提高客户端的并发能力？下面做了一组测试，分别使用同步、异步、协程+同步等方式请求百度的首页，运行10s，对比一下QPS。

测试环境

2017年MBP，2核4线程，16GB内存。

Maven依赖

<properties>
    <kotlin.version>1.3.11</kotlin.version>
</properties>

<dependency>
    <groupId>org.jetbrains.kotlin</groupId>
    <artifactId>kotlin-stdlib-jdk8</artifactId>
    <version>${kotlin.version}</version>
</dependency>

<dependency>
    <groupId>org.jetbrains.kotlin</groupId>
    <artifactId>kotlin-test</artifactId>
    <version>${kotlin.version}</version>
    <scope>test</scope>
</dependency>

<dependency>
    <groupId>org.jetbrains.kotlinx</groupId>
    <artifactId>kotlinx-coroutines-core</artifactId>
    <version>1.1.0</version>
</dependency>

<dependency>
    <groupId>org.apache.httpcomponents</groupId>
    <artifactId>httpasyncclient</artifactId>
    <version>4.1</version>
    <scope>test</scope>
</dependency>

<dependency>
    <groupId>khttp</groupId>
    <artifactId>khttp</artifactId>
    <version>0.1.0</version>
</dependency>

<dependency>
    <groupId>io.github.rybalkinsd</groupId>
    <artifactId>kohttp</artifactId>
    <version>0.5.0</version>
</dependency>

<dependency>
    <groupId>com.squareup.okhttp3</groupId>
    <artifactId>okhttp</artifactId>
    <version>3.12.1</version>
</dependency>

Kotlin多线程同步

使用Kotlin的khttp发起同步请求。

import org.junit.Test
import java.util.concurrent.atomic.AtomicInteger

public class TestHttpSyncClient {

    val api = "https://www.baidu.com"

    val completedCount = AtomicInteger(0)
    var failedCount = AtomicInteger(0)

    @Test
    fun test() {

        for (i in 0..80) {

            Thread() {

                while (true) {

                    try {

                        var response = khttp.get(api)

                        if (response.statusCode == 200) {
                            completedCount.incrementAndGet()
                        } else {
                            failedCount.incrementAndGet()
                        }
                    }catch (e : Exception) {

                        failedCount.incrementAndGet()
                    }
                }
            }.start()
        }

        Thread.sleep(10 * 1000)
        println("completedCount: ${completedCount}, failedCount: ${failedCount}");
        System.exit(0);
    }
}

线程开到80个，QPS可达2600左右。

completedCount: 26089, failedCount: 111

HttpAsyncClients异步

使用HttpAsyncClients发起异步请求。

import org.apache.http.HttpResponse;
import org.apache.http.client.methods.HttpGet;
import org.apache.http.concurrent.FutureCallback;
import org.apache.http.impl.nio.client.CloseableHttpAsyncClient;
import org.apache.http.impl.nio.client.HttpAsyncClients;
import org.junit.Test;

import java.util.concurrent.atomic.AtomicInteger;

public class TestHttpAsyncClient implements Runnable {

   public final static AtomicInteger completedCount = new AtomicInteger(0);
   public final static AtomicInteger failedCount = new AtomicInteger(0);
   public final static AtomicInteger canceledCount = new AtomicInteger(0);

   @Test
   public void test() {

       //启动线程
       TestHttpAsyncClient testHttpAsyncClient = new TestHttpAsyncClient();
       new Thread(testHttpAsyncClient).start();

       //测试10s
       try {

           Thread.sleep(10 * 1000);
           System.out.println(String.format("completedCount: %d, failedCount: %d, canceledCount: %d",
                   completedCount.get(), failedCount.get(), canceledCount.get()));

           System.exit(0);
       } catch (Exception e) {

           System.err.println(String.format("System.exit exception: %s", e));
       }
   }

   public void run() {

       CloseableHttpAsyncClient httpclient = HttpAsyncClients.custom()
               .setMaxConnTotal(128)
               .setMaxConnPerRoute(128)
               .build();
       httpclient.start();

       final HttpGet request = new HttpGet("https://www.baidu.com");

       while (true) {

           httpclient.execute(request, new FutureCallback<HttpResponse>() {

               @Override
               public void completed(HttpResponse httpResponse) {

                   if (httpResponse.getStatusLine().getStatusCode() == 200) {
                       completedCount.incrementAndGet();
                   }else {
                       failedCount.incrementAndGet();
                 }
               }

               @Override
               public void failed(Exception e) {
                   failedCount.incrementAndGet();
               }

               @Override
               public void cancelled() {
                   canceledCount.incrementAndGet();
               }
           });
       }
   }
}

并发设置为128，QPS可以达到2000左右。

completedCount: 19319, failedCount: 125, canceledCount: 0

通过调试可以发现，只需要4个dispatcher线程，就可以满足需要。

image.png

Kotlin协程+khttp

使用Kotlin Coroutine+khttp同步请求测试。

import khttp.responses.Response
import kotlinx.coroutines.Deferred
import kotlinx.coroutines.GlobalScope
import kotlinx.coroutines.async
import kotlinx.coroutines.launch
import java.util.concurrent.atomic.AtomicInteger
import org.junit.Test

public class TestKotlinCoroutine {

    val api = "https://www.baidu.com"
    val completedCount = AtomicInteger(0)
    val failedCount = AtomicInteger(0)

    @Test
    fun test() {

        Thread(){

            while (true) {

                val resp =
                        GlobalScope.launch {

                            val result: Deferred<Response> = async {

                                get()
                            }

                            result.await()
                        }
            }
        }.start()

        Thread.sleep(10 * 1000)
        println("completedCount: ${completedCount}, failedCount: ${failedCount}");
        System.exit(0);
    }

    suspend fun get(): khttp.responses.Response {

        var response = khttp.get(api)

        if (response.statusCode == 200) {
            completedCount.addAndGet(1)
        }else {
            failedCount.addAndGet(1)
        }

        return response;
    }
}

协程版本的性能甚至不如同步版本，QPS只有50。换一种方式，起多个协程，每个线程里面循环请求，QPS也一样只有50左右。

completedCount: 498, failedCount: 0

这个版本性能很差，通过调试可以发现，最多也就启动了4个worker线程，并且khttp又不支持协程阻塞。

image.png

下面是一个优化的版本，将工作放到Dispatchers.IO里面去做，尽量多起一些worker线程。

public class TestKotlinCoroutineClient {

    val api = "https://www.baidu.com"

    val completedCount = AtomicInteger(0)
    var failedCount = AtomicInteger(0)
    val time = 10000

    @Test
    fun test() = runBlocking {

        val start = System.currentTimeMillis()
        val channel = Channel<Int>()

        repeat(time) {
            launch { channel.send(get()) }
        }

        repeat(time) {
            val code = channel.receive()
            if (code == 200) {
                completedCount.incrementAndGet()
            } else {
                failedCount.incrementAndGet()
            }
        }

        val end = System.currentTimeMillis()
        println("completedCount: ${completedCount}, failedCount: ${failedCount}, cost${(end - start) / 1000}");
        System.exit(0);
    }

    suspend fun get() = withContext(Dispatchers.IO) { khttp.get(api) }.statusCode
}

性能可以提升到1200左右。调试可以发现线程数量最多可达68个。

image.png

Kotlin协程+kohttp

kohttp看起来支持协程。使用kohttp请求一万次，花费时间100s左右，QPS为100。

import io.github.rybalkinsd.kohttp.ext.asyncHttpGet
import junit.framework.Assert.assertEquals
import kotlinx.coroutines.runBlocking
import org.junit.Test
import kotlin.system.measureTimeMillis

public class TestKotlinCoroutine {

    val api = "https://www.baidu.com"

    @Test
    fun `many async invokes of httpGet`() {
        measureTimeMillis {
            GlobalScope.launch {
                val tasks = List(10000) {
                    api.asyncHttpGet()
                }
                tasks.map { r ->
                    r.await().also { it.close() }
                }.forEach {
                    assertEquals(200, it.code())
                }
            }
        }.also { println("$it ms") }
    }
}

分析一下kohttp的实现，可以发现它只是封装了okhttp3，使用了异步模式。okhttp3本身不支持协程，所以性能也不会太好。

fun String.asyncHttpGet(client: Call.Factory = defaultHttpClient): Deferred<Response> =
    GlobalScope.async(context = Unconfined) {
        client.suspendCall(Request.Builder().url(this@asyncHttpGet).build())
    }

internal suspend fun Call.Factory.suspendCall(request: Request): Response =
    suspendCoroutine { cont ->
        newCall(request).enqueue(object : Callback {
            override fun onResponse(call: Call, response: Response) {
                cont.resume(response)
            }

            override fun onFailure(call: Call, e: IOException) {
                cont.resumeWithException(e)
            }
        })
    }

Go协程

顺便试一下Go协程的性能。Go协程版本在使用keepalive的情况下，256个协程的QPS可以达到1700左右。注意需要读一下resp.Body，这样keepalive才会生效。参考文章：A brief intro of TCP keep-alive in Go’s HTTP implementation。

package main

import (
    "fmt"
    "io"
    "io/ioutil"
    "net/http"
    "os"
    "runtime"
    "sync"
    "sync/atomic"
    "time"
)

var completedCount uint64
var failedCount uint64
var lock = &sync.Mutex{}

func main() {

    runtime.GOMAXPROCS(4)
    for i := 0; i < 256; i++ {
        clt := newQPSClient()
        go clt.test()
    }

    time.Sleep(10 * time.Second)
    fmt.Printf("completedCount: %d, failedCount: %d\n", completedCount, failedCount)
    os.Exit(0)
}

type QPSClient struct {
    clt *http.Client
}

func newQPSClient() *QPSClient {
    return &QPSClient{
        clt: &http.Client{},
    }
}

func (qc *QPSClient) test() {

    for {
        resp, err := qc.clt.Get("https://www.baidu.com")
        if err == nil && (resp.StatusCode == 200) {

            _, err = io.Copy(ioutil.Discard, resp.Body)
            resp.Body.Close()
            atomic.AddUint64(&completedCount, 1)
        } else {

            _, err = io.Copy(ioutil.Discard, resp.Body)
            atomic.AddUint64(&failedCount, 1)
        }
    }
}

结论

请求方式	QPS
Kotlin多线程同步	80个线程，QPS 2600
HttpAsyncClients异步	最大128个连接，QPS 2000
Kotlin协程+khttp	循环起协程+khttp同步，QPS 50
Kotlin协程+kohttp	循环起协程+okhttp3异步，QPS 100
Go协程	256个协程，QPS 1700

上面这些测试，缺乏更精细的设置，比如HttpAsyncClients 128个连接使用了多少个线程？Kotlin和Go的协程版本使用了多少线程？还有没有提升空间？

Coroutine运行有很多受限条件，不能堵塞在操作系统会堵塞线程的地方，需要自己实现对这些堵塞API的处理，在用户态做上下文的保存和恢复。Java生态圈中缺乏对协程支持良好的基础库，导致不能发挥协程真正的威力，使用异步IO是更好的解决办法。