前言

surging 对外沉寂了一段时间了,但是作者并没有闲着,而是针对于客户的需要添加了不少功能,也给我带来了不少外快收益, 就比如协议转化,consul 的watcher 机制,JAVA版本,skywalking 升级支持8.0,.升级NET 6.0 ,而客户自己扩展支持服务编排流程引擎,后期客户还需要扩展定制coap ,XMPP等协议。而今天写这篇文章的目的针对于修改基于netty 的异步非阻塞业务逻辑操作

问题描述

年前客户把JAVA版本进行了测试,产生了不少问题,客户也比较茫然,因为有内存泄漏,通过jmeter压测,并发始终上不来,通过半个月的努力,终于把问题解决了,预估JAVA版本并发能达到2万左右,以下是客户通过设置jmeter压测实例

解决方案

当客户把问题抛给我后,我第一反应是IO线程被阻塞造成的,而这样就可以把问题定位在netty 的处理上,而处理server 端代码是NettyServerMessageListener,而其中ServerHandler的channelRead是处理业务逻辑的,在这当中我是通过ThreadPoolExecutor执行异步处理,可以看看NettyServerMessageListener代码:

public class NettyServerMessageListener implements IMessageListener {
private Thread thread;
private static final Logger logger = LoggerFactory.getLogger(NettyServerMessageListener.class);
private ChannelFuture channel;
private final ITransportMessageDecoder transportMessageDecoder;
private final ITransportMessageEncoder transportMessageEncoder;
ReceivedDelegate Received = new ReceivedDelegate();
@Inject
public NettyServerMessageListener( ITransportMessageCodecFactory codecFactory)
{
this.transportMessageEncoder = codecFactory.GetEncoder();
this.transportMessageDecoder = codecFactory.GetDecoder();
} public void StartAsync(final String serverAddress) {
thread = new Thread(new Runnable() {
int parallel = Runtime.getRuntime().availableProcessors();
final DefaultEventLoopGroup eventExecutors = new DefaultEventLoopGroup(parallel);
ThreadFactory threadFactory = new DefaultThreadFactory("rpc-netty", true);
public void run() {
String[] array = serverAddress.split(":");
logger.debug("准备启动服务主机,监听地址:" + array[0] + "" + array[1] + "。");
EventLoopGroup bossGroup = new NioEventLoopGroup();
EventLoopGroup workerGroup = new NioEventLoopGroup(parallel,threadFactory);
ServerBootstrap bootstrap = new ServerBootstrap();
bootstrap.group(bossGroup, workerGroup).option(ChannelOption.SO_BACKLOG,128)
.childOption(ChannelOption.SO_KEEPALIVE,true).childOption(ChannelOption.TCP_NODELAY, true).channel(NioServerSocketChannel.class)
.childHandler(new ChannelInitializer<NioSocketChannel>() {
@Override
protected void initChannel(NioSocketChannel socketChannel) throws Exception {
socketChannel.pipeline()
.addLast(new LengthFieldPrepender(4))
.addLast(new LengthFieldBasedFrameDecoder(Integer.MAX_VALUE, 0, 4, 0, 4))
.addLast(new ServerHandler(eventExecutors,new ReadAction<ChannelHandlerContext, TransportMessage>() {
@Override
public void run() {
IMessageSender sender = new NettyServerMessageSender(transportMessageEncoder, this.parameter);
onReceived(sender, this.parameter1);
}
},transportMessageDecoder)
);
}
})
.option(ChannelOption.SO_BACKLOG, 128)
.childOption(ChannelOption.ALLOCATOR, PooledByteBufAllocator.DEFAULT);
try { String host = array[0];
int port = Integer.parseInt(array[1]);
channel = bootstrap.bind(host, port).sync();
logger.debug("服务主机启动成功,监听地址:" + serverAddress + "。");
} catch (Exception e) {
if (e instanceof InterruptedException) {
logger.info("Rpc server remoting server stop");
} else {
logger.error("Rpc server remoting server error", e);
} }
}
});
thread.start();
} @Override
public ReceivedDelegate getReceived() {
return Received;
} public void onReceived(IMessageSender sender, TransportMessage message) {
if (Received == null)
return;
Received.notifyX(sender,message);
} private class ReadAction<T,T1> implements Runnable
{
public T parameter;
public T1 parameter1;
public void setParameter( T tParameter,T1 tParameter1) {
parameter = tParameter;
parameter1 = tParameter1;
} @Override
public void run() { }
} private class ServerHandler extends ChannelInboundHandlerAdapter {
private final DefaultEventLoopGroup serverHandlerPool;
private final ReadAction<ChannelHandlerContext, TransportMessage> serverRunnable;
private final ITransportMessageDecoder transportMessageDecoder; public ServerHandler(final DefaultEventLoopGroup threadPoolExecutor, ReadAction<ChannelHandlerContext, TransportMessage> runnable,
ITransportMessageDecoder transportMessageDecoder) {
this.serverHandlerPool = threadPoolExecutor;
this.serverRunnable = runnable;
this.transportMessageDecoder = transportMessageDecoder;
} @Override
public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause) {
logger.warn("与服务器:" + ctx.channel().remoteAddress() + "通信时发送了错误。");
ctx.close();
} @Override
public void channelReadComplete(ChannelHandlerContext context) {
context.flush();
} @Override
public void channelRead(ChannelHandlerContext channelHandlerContext, Object message) throws Exception {
ByteBuf buffer = (ByteBuf) message;
try {
byte[] data = new byte[buffer.readableBytes()];
buffer.readBytes(data);
serverHandlerPool.execute(() -> {
TransportMessage transportMessage = null;
try {
transportMessage = transportMessageDecoder.Decode(data);
} catch (IOException e) {
e.printStackTrace();
}
serverRunnable.setParameter(channelHandlerContext, transportMessage);
serverRunnable.run();
});
}
finally {
ReferenceCountUtil.release(message);
}
}
}
}

ThreadPoolExecutor代码:

    public static ThreadPoolExecutor makeServerThreadPool(final String serviceName, int corePoolSize, int maxPoolSize) {
ThreadPoolExecutor serverHandlerPool = new ThreadPoolExecutor(
corePoolSize,
maxPoolSize,
60L,
TimeUnit.SECONDS,
new ArrayBlockingQueue<Runnable>( 10000));
/*
new LinkedBlockingQueue<Runnable>(10000),
r -> new Thread(r, "netty-rpc-" + serviceName + "-" + r.hashCode()),
new ThreadPoolExecutor.AbortPolicy());*/ return serverHandlerPool;
}

后面通过查找官方的文档发现以下addLast是IO线程阻塞调用

 .addLast(new ServerHandler(eventExecutors,new ReadAction<ChannelHandlerContext, TransportMessage>() {
@Override
public void run() {
IMessageSender sender = new NettyServerMessageSender(transportMessageEncoder, this.parameter);
onReceived(sender, this.parameter1);
}
},transportMessageDecoder)

后面通过使用EventExecutorGroup把IO线程与业务线程进行分离,把耗时业务处理添加到EventExecutorGroup进行处理,首先EventExecutorGroup代码如下

    public static final EventExecutorGroup execThreadPool = new DefaultEventExecutorGroup( Runtime.getRuntime().availableProcessors()*2,
(ThreadFactory) r -> {
Thread thread = new Thread(r);
thread.setName("custom-tcp-exec-"+r.hashCode());
return thread;
},
100000,
RejectedExecutionHandlers.reject()
);

而addLast的ServerHandler添加了EventExecutorGroup, 最新的NettyServerMessageListener代码如下:

public class NettyServerMessageListener implements IMessageListener {
private Thread thread;
private static final Logger logger = LoggerFactory.getLogger(NettyServerMessageListener.class);
private ChannelFuture channel;
private final ITransportMessageDecoder transportMessageDecoder;
private final ITransportMessageEncoder transportMessageEncoder;
ReceivedDelegate Received = new ReceivedDelegate();
@Inject
public NettyServerMessageListener( ITransportMessageCodecFactory codecFactory)
{
this.transportMessageEncoder = codecFactory.GetEncoder();
this.transportMessageDecoder = codecFactory.GetDecoder();
} public void StartAsync(final String serverAddress) {
thread = new Thread(new Runnable() {
public void run() {
String[] array = serverAddress.split(":");
logger.debug("准备启动服务主机,监听地址:" + array[0] + "" + array[1] + "。");
EventLoopGroup bossGroup = new NioEventLoopGroup(1);
EventLoopGroup workerGroup = new NioEventLoopGroup();
ServerBootstrap bootstrap = new ServerBootstrap();
bootstrap.group(bossGroup, workerGroup).channel(NioServerSocketChannel.class)
.childHandler(new ChannelInitializer<NioSocketChannel>() {
@Override
protected void initChannel(NioSocketChannel socketChannel) throws Exception {
socketChannel.pipeline()
.addLast(new LengthFieldPrepender(4))
.addLast(new LengthFieldBasedFrameDecoder(Integer.MAX_VALUE, 0, 4, 0, 4))
.addLast(ThreadPoolUtil.execThreadPool, "handler",new ServerHandler(new ReadAction<ChannelHandlerContext, TransportMessage>() {
@Override
public void run() {
IMessageSender sender = new NettyServerMessageSender(transportMessageEncoder, this.parameter);
onReceived(sender, this.parameter1);
}
},transportMessageDecoder)
);
}
})
.option(ChannelOption.SO_BACKLOG, 128)
.childOption(ChannelOption.ALLOCATOR, PooledByteBufAllocator.DEFAULT);
try { String host = array[0];
int port = Integer.parseInt(array[1]);
channel = bootstrap.bind(host, port).sync();
logger.debug("服务主机启动成功,监听地址:" + serverAddress + "。");
} catch (Exception e) {
if (e instanceof InterruptedException) {
logger.info("Rpc server remoting server stop");
} else {
logger.error("Rpc server remoting server error", e);
} }
}
});
thread.start();
} @Override
public ReceivedDelegate getReceived() {
return Received;
} public void onReceived(IMessageSender sender, TransportMessage message) {
if (Received == null)
return;
Received.notifyX(sender,message);
} private class ReadAction<T,T1> implements Runnable
{
public T parameter;
public T1 parameter1;
public void setParameter( T tParameter,T1 tParameter1) {
parameter = tParameter;
parameter1 = tParameter1;
} @Override
public void run() { }
} private class ServerHandler extends ChannelInboundHandlerAdapter { private final ReadAction<ChannelHandlerContext, TransportMessage> serverRunnable;
private final ITransportMessageDecoder transportMessageDecoder; public ServerHandler(ReadAction<ChannelHandlerContext, TransportMessage> runnable,
ITransportMessageDecoder transportMessageDecoder) { this.serverRunnable = runnable;
this.transportMessageDecoder = transportMessageDecoder;
} @Override
public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause) {
logger.warn("与服务器:" + ctx.channel().remoteAddress() + "通信时发送了错误。");
ctx.close();
} @Override
public void channelReadComplete(ChannelHandlerContext context) {
context.flush();
} @Override
public void channelRead(ChannelHandlerContext channelHandlerContext, Object message) throws Exception {
ByteBuf buffer = (ByteBuf) message;
try {
byte[] data = new byte[buffer.readableBytes()];
buffer.readBytes(data);
TransportMessage transportMessage = transportMessageDecoder.Decode(data);
serverRunnable.setParameter(channelHandlerContext, transportMessage);
serverRunnable.run();
}
finally {
ReferenceCountUtil.release(message);
}
}
}
}

通过以上修改,再通过jmeter压测已经不会出现timeout 问题,就连stage 网关-》.NET微服务-》JAVA微服务都没有Time out问题产生,jmeter的user thread拉长到2000也没有出现问题。

通过以上思路把.NET版本的surging 社区版本也进行了修改,已经提交到github,首先把ServiceHost中的serverMessageListener.Received 中的Task.Run移除,ServerHandler中ChannelRead进行移除,然后addLast的ServerHandler添加了EventExecutorGroup.通过以上修改再通过压测发现可以支持20万+ ,也未发现内存泄漏问题,执行client 1万次 ,服务端cpu 在6%左右,响应速度在1.1秒左右,可以开启多个surging 的client 进行压测,cpu 会叠加上升,响应速度没有影响,以下是执行1万次压测

surging 社区版本开源地址

总结

通过5年研发,surging 从原来的最初的基于netty 的RPC发展到现在可以支持多协议,多语言的异构微服务引擎,不仅是技术的提高,也带来名利的收益,只要不断坚持,终究能看到成果,我也会一直更新,为企业和社区用户带来自己的绵薄之力,让企业能更好的掌握微服务解决方案,已解决现在行业各种不同的业务需求。

suging闲谈-netty 的异步非阻塞IO线程与业务线程分离的更多相关文章

  1. 异步非阻塞IO的Python Web框架--Tornado

    Tornado的全称是Torado Web Server,从名字上就可知它可用作Web服务器,但同时它也是一个Python Web的开发框架.最初是在FriendFeed公司的网站上使用,FaceBo ...

  2. 转一贴,今天实在写累了,也看累了--【Python异步非阻塞IO多路复用Select/Poll/Epoll使用】

    下面这篇,原理理解了, 再结合 这一周来的心得体会,整个框架就差不多了... http://www.haiyun.me/archives/1056.html 有许多封装好的异步非阻塞IO多路复用框架, ...

  3. Python异步非阻塞IO多路复用Select/Poll/Epoll使用,线程,进程,协程

    1.使用select模拟socketserver伪并发处理客户端请求,代码如下: import socket import select sk = socket.socket() sk.bind((' ...

  4. nodejs的异步非阻塞IO

    简单表述一下:发启向系统IO操作请求,系统使用线程池IO操作,执行完放到事件队列里,node主线程轮询事件队列,读取结果与调用回调.所以说node并非真的单线程,还是使用了线程池的多线程. 上个图看看 ...

  5. swoole与php协程实现异步非阻塞IO开发

    “协程可以在遇到阻塞的时候中断主动让渡资源,调度程序选择其他的协程运行.从而实现非阻塞IO” 然而php是不支持原生协程的,遇到阻塞时如不交由异步进程来执行是没有任何意义的,代码还是同步执行的,如下所 ...

  6. [Flask] 异步非阻塞IO实现

    Flask默认是不支持非阻塞IO的,表现为: 当 请求1未完成之前,请求2是需要等待处理状态,效率非常低. 在flask中非阻塞实现可以由2种: 启用flask多线程机制 # Flask from f ...

  7. 谈谈对不同I/O模型的理解 (阻塞/非阻塞IO,同步/异步IO)

    一.关于I/O模型的问题 最近通过对ucore操作系统的学习,让我打开了操作系统内核这一黑盒子,与之前所学知识结合起来,解答了长久以来困扰我的关于I/O的一些问题. 1. 为什么redis能以单工作线 ...

  8. 同步异步阻塞非阻塞Reactor模式和Proactor模式 (目前JAVA的NIO就属于同步非阻塞IO)

    在高性能的I/O设计中,有两个比较著名的模式Reactor和Proactor模式,其中Reactor模式用于同步I/O,而Proactor运用于异步I/O操作. 在比较这两个模式之前,我们首先的搞明白 ...

  9. 【面试】详解同步/异步/阻塞/非阻塞/IO含义与案例

    本文详解同步.异步.阻塞.非阻塞,以及IO与这四者的关联,毕竟我当初刚认识这几个名词的时候也是一脸懵. 目录 1.同步阻塞.同步非阻塞.异步阻塞.异步非阻塞 1.同步 2.异步 3.阻塞 4.非阻塞 ...

随机推荐

  1. c# - 常量定义与赋值

    1.前言 c#与Java很相似,但是不一样,又与js(JavaScript)相似,但是也不一样,所以我认为c#是Java和 js的孩子. 2.常量定义 字符串: const string = &quo ...

  2. centos7 单用户模式修改root密码

    1. 在虚拟机重启客户机后.会出现下面进入界面.按e键 2.按了e键后,会出现下面的界面.此时按↓键.找到linux16 3.将光标移动到UTF-8后面,添加init=/bin/sh,并按 ctrl  ...

  3. 第10组 Alpha冲刺 (1/6)

    1.1基本情况 ·队名:今晚不睡觉 ·组长博客:https://www.cnblogs.com/cpandbb/ ·作业博客:https://edu.cnblogs.com/campus/fzu/FZ ...

  4. 关于less使用初学者常犯错误

    1.多层嵌套 2.定义变量复用性不高 3.方法定义多余参数.

  5. HIVE优化学习笔记

    概述 之前写过关于hive的已经有两篇随笔了,但是作者依然还是一枚小白,现在把那些杂七杂八的总结一下,供以后查阅和总结.今天的文章介绍一下hive的优化.hive是好多公司都在使用的东西,也有好多大公 ...

  6. 浅解XXE与Portswigger Web Sec

    XXE与Portswigger Web Sec ​ 相关链接: ​ 博客园 ​ 安全脉搏 ​ FreeBuf 简介XML XML,可扩展标记语言,标准通用标记语言的子集.XML的简单易于在任何应用程序 ...

  7. leetcode 206. 反转链表 及 92. 反转链表 II

    206. 反转链表 问题描述 反转一个单链表. 示例: 输入: 1->2->3->4->5->NULL 输出: 5->4->3->2->1-> ...

  8. go语言 strconv.ParseInt 的实现分析

    字符串与数值之间进行转换是一个高频操作,在go语言中,SDK提供 strconv.ParseInt 将字符串转换为数值,strconv.FormatInt 可以将数值转换为字符串. 1.首先看下 st ...

  9. 毫米转像素dpi

    public static double MillimeterToPixel_X(double length) //length是毫米,1厘米=10毫米 { System.Windows.Forms. ...

  10. C++构造函数语义学(一)(基于C++对象模型)

    如果一个类没有自己的构造函数,编译器会在需要的时候为其合成一个出来,俗称:合成默认构造函数.但是请注意是在需要的时候,并不是所有情况. 请看下面代码: 1 #include<iostream&g ...