Zuul 2相对zuul 1 由同步改进为异步机制,没有了同步阻塞,全部基于事件驱动模型编程,线程模型也变得简单。
zuul做为一个网关接受客户端的请求--服务端,又要和后端的服务建立链接,把请求转发给后端服务--客户端。下面我们来分析zuul 是怎么通过一个线程池来实现的,即服务端和客户端用同一个线程池,通过netty 编程这很容易,空口无凭,我们来看zuul 是怎么实现的。
线程池初始化
线程池的个数如下
public DefaultEventLoopConfig()
{
eventLoopCount = WORKER_THREADS.get() > 0 ? WORKER_THREADS.get() : PROCESSOR_COUNT;
acceptorCount = ACCEPTOR_THREADS.get();
}
boss 线程数是1,
worker 线程数是:PROCESSOR_COUNT 缺省值如下,为cpu的核数。
zuul 异步线程模型,吞吐量很高,所以线程的个数基本按cpu核数来,这样上下文切换的开销很少。
private static final int PROCESSOR_COUNT = Runtime.getRuntime().availableProcessors();
Zuul server 的start 方法入手
public void start(boolean sync)
{
//ServerGroup 是zuul 对netty eventloop的简单封装,
serverGroup = new ServerGroup("Salamander", eventLoopConfig.acceptorCount(), eventLoopConfig.eventLoopCount(), eventLoopGroupMetrics);
//使用epoll还是机遇jdk的多路复用select来实现事件驱动,epoll 只支持Linux,使用
serverGroup.initializeTransport();
try {
List allBindFutures = new ArrayList<>();
// Setup each of the channel initializers on requested ports.
for (Map.Entry entry : portsToChannelInitializers.entrySet())
{
allBindFutures.add(setupServerBootstrap(entry.getKey(), entry.getValue()));
}
// Once all server bootstraps are successfully initialized, then bind to each port.
for (ChannelFuture f: allBindFutures) {
// Wait until the server socket is closed.
ChannelFuture cf = f.channel().closeFuture();
if (sync) {
cf.sync();
}
}
}
catch (InterruptedException e) {
Thread.currentThread().interrupt();
}
}
Epoll 还是Selector
我们知道,操作系统底层的IO方式有select,poll,epoll,最早的是select 机制,但是select机制对文件句柄的个数有限制,而且需要迭代,而epoll 无限制,而且不需要迭代,效率最高,jdk也有epoll 的api,不过默认是采用水平出发的方式,netty 的epoll 是采用边缘触发方式,效率更高,所以写netty的都会做个适配,是用epoll还是用select。
服务端
private ChannelFuture setupServerBootstrap(int port, ChannelInitializer channelInitializer)
throws InterruptedException
{
//设置上面创建的两个线程池
ServerBootstrap serverBootstrap = new ServerBootstrap().group(
serverGroup.clientToProxyBossPool,
serverGroup.clientToProxyWorkerPool);
// Choose socket options.
Map channelOptions = new HashMap<>();
channelOptions.put(ChannelOption.SO_BACKLOG, 128);
//channelOptions.put(ChannelOption.SO_TIMEOUT, SERVER_SOCKET_TIMEOUT.get());
channelOptions.put(ChannelOption.SO_LINGER, -1);
channelOptions.put(ChannelOption.TCP_NODELAY, true);
channelOptions.put(ChannelOption.SO_KEEPALIVE, true);
// Choose EPoll or NIO.
if (USE_EPOLL.get()) {
LOG.warn("Proxy listening with TCP transport using EPOLL");
serverBootstrap = serverBootstrap.channel(EpollServerSocketChannel.class);
channelOptions.put(EpollChannelOption.TCP_DEFER_ACCEPT, Integer.valueOf(-1));
}
else {
LOG.warn("Proxy listening with TCP transport using NIO");
serverBootstrap = serverBootstrap.channel(NioServerSocketChannel.class);
}
// Apply socket options.
for (Map.Entry optionEntry : channelOptions.entrySet()) {
serverBootstrap = serverBootstrap.option(optionEntry.getKey(), optionEntry.getValue());
}
//设置channelInitializer,后面分析的入口就在这里了。
serverBootstrap.childHandler(channelInitializer);
serverBootstrap.validate();
LOG.info("Binding to port: " + port);
// Flag status as UP just before binding to the port.
serverStatusManager.localStatus(InstanceInfo.InstanceStatus.UP);
// Bind and start to accept incoming connections.
return serverBootstrap.bind(port).sync();
}
客户端
服务端接收到请求后,请求经过一序列的filter 处理,会交给zuul的ProxyEndpoint 来把请求转发给后端服务,ProxyEndpoint这里 需要做路由和对后端链接的建立,即实现netty的客户端,执行的入口如下:
ProxyEndpoint 的 proxyRequestToOrigin 方法:
attemptNum += 1;
requestStat = createRequestStat();
origin.preRequestChecks(zuulRequest);
concurrentReqCount++;
//关键是这里,channelCtx.channel().eventLoop()
promise = origin.connectToOrigin(zuulRequest, channelCtx.channel().eventLoop(), attemptNum, passport, chosenServer);
logOriginServerIpAddr();
currentRequestAttempt = origin.newRequestAttempt(chosenServer.get(), context, attemptNum);
requestAttempts.add(currentRequestAttempt);
passport.add(PassportState.ORIGIN_CONN_ACQUIRE_START);
if (promise.isDone()) {
operationComplete(promise);
} else {
promise.addListener(this);
}
}
上面的这行代码中的channelCtx.channel().eventLoop(),就是当前netty 接入端的worker event loop。
promise = origin.connectToOrigin(zuulRequest, channelCtx.channel().eventLoop(), attemptNum, passport, chosenServer);
这里主要是实现如下几点:
- 负载均衡,选择一台机器。
- 为对应的机器创建链接池。
- 从链接池活着链接。
第一次建立链接时,最终会执行如下的代码:
public ChannelFuture connect(final EventLoop eventLoop, String host, final int port, CurrentPassport passport) {
Class socketChannelClass;
if (Server.USE_EPOLL.get()) {
socketChannelClass = EpollSocketChannel.class;
} else {
socketChannelClass = NioSocketChannel.class;
}
final Bootstrap bootstrap = new Bootstrap()
.channel(socketChannelClass)
.handler(channelInitializer)
.group(eventLoop)
.attr(CurrentPassport.CHANNEL_ATTR, passport)
.option(ChannelOption.CONNECT_TIMEOUT_MILLIS, connPoolConfig.getConnectTimeout())
.option(ChannelOption.SO_KEEPALIVE, connPoolConfig.getTcpKeepAlive())
.option(ChannelOption.TCP_NODELAY, connPoolConfig.getTcpNoDelay())
.option(ChannelOption.SO_SNDBUF, connPoolConfig.getTcpSendBufferSize())
.option(ChannelOption.SO_RCVBUF, connPoolConfig.getTcpReceiveBufferSize())
.option(ChannelOption.WRITE_BUFFER_HIGH_WATER_MARK, connPoolConfig.getNettyWriteBufferHighWaterMark())
.option(ChannelOption.WRITE_BUFFER_LOW_WATER_MARK, connPoolConfig.getNettyWriteBufferLowWaterMark())
.option(ChannelOption.AUTO_READ, connPoolConfig.getNettyAutoRead())
.remoteAddress(new InetSocketAddress(host, port));
return bootstrap.connect();
}
上面的代码是不是很熟悉,是netty客户端的实现,绑定的eventloop 就是前面传递进来的即接入端的eventLoop,这样zuul 就是实现了接入端的io 操作和后端服务的读写都是绑定到同一个eventLoop线程上。
总结
zuul2的线程模型好简单,就是netty的一个eventloop 线程池,来处理所有的请求和远程调用。
从上面的图可以看出,一个请求进来和调用远程服务,以及回写都是由同一个线程来完成的,完全没有上下文切换。zuul2 用一个线程池搞定所有的这些,这都是得益于netty 异步编程的威力。