最近在写一个后台中间件的原型,主要是做消息的分发和透传。因为要用Java实现,所以网络通信框架的第一选择当然就是Netty了,使用的是Netty 4版本。Netty果然效率很高,不用做太多努力就能达到一个比较高的tps。但使用过程中也碰到了一些问题,个人觉得都是比较经典而在网上又不太容易查找到相关资料的问题,所以在此总结一下。
1.Context Switch过高
压测时用nmon监控内核,发现Context Switch高达30w+。这明显不正常,但JVM能有什么导致Context Switch。参考之前整理过的恐龙书《Operating System Concept》的读书笔记《进程调度》和Wiki上的Context Switch介绍,进程/线程发生上下文切换的原因有:
- I/O等待:在多任务系统中,进程主动发起I/O请求,但I/O设备还没有准备好,所以会发生I/O阻塞,进程进入Wait状态。
- 时间片耗尽:在多任务分时系统中,内核分配给进程的时间片已经耗尽了,进程进入Ready状态,等待内核重新分配时间片后的执行机会。
- 硬件中断:在抢占式的多任务分时系统中,I/O设备可以在任意时刻发生中断,CPU会停下当前正在执行的进程去处理中断,因此进程进入Ready状态。
根据分析,重点就放在第一个和第二个因素上。
进程与线程的上下文切换
之前的读书笔记里总结的是进程的上下文切换原因,那线程的上下文切换又有什么不同呢?在StackOverflow上果然找到了提问thread context switch vs process context switch:
“The main distinction between a thread switch and a process switch is that during a thread switch, the virtual memory space remains the same, while it does not during a process switch. Both types involve handing control over to the operating system kernel to perform the context switch. The process of switching in and out of the OS kernel along with the cost of switching out the registers is the largest fixed cost of performing a context switch.
A more fuzzy cost is that a context switch messes with the processors cacheing mechanisms. Basically, when you context switch, all of the memory addresses that the processor “remembers” in it’s cache effectively become useless. The one big distinction here is that when you change virtual memory spaces, the processor’s Translation Lookaside Buffer (TLB) or equivalent gets flushed making memory accesses much more expensive for a while. This does not happen during a thread switch.”通过排名第一的大牛的解答了解到,进程和线程的上下文切换都涉及进出系统内核和寄存器的保存和还原,这是它们的最大开销。但与进程的上下文切换相比,线程还是要轻量一些,最大的区别是线程上下文切换时虚拟内存地址保持不变,所以像TLB等CPU缓存不会失效。但要注意的是另一份提问What is the overhead of a context-switch?的中提到了:Intel和AMD在2008年引入的技术可能会使TLB不失效。感兴趣的话请自行研究吧。
1.1 非阻塞I/O
针对第一个因素I/O等待,最直接的解决办法就是使用非阻塞I/O操作。在Netty中,就是服务端和客户端都使用NIO。
这里在说一下如何主动的向Netty的Channel写入数据,因为网络上搜到的资料都是千篇一律:服务端就是接到请求后在Handler中写入返回数据,而客户端的例子竟然也都是在Handler里Channel Active之后发送数据。因为要做消息透传,而且是向下游系统发消息时是异步非阻塞的,网上那种例子根本没法用,所以在这里说一下我的方法吧。
关于服务端,在接收到请求后,在channelRead0()中通过ctx.channel()得到Channel,然后就通过ThreadLocal变量或其他方法,只要能把这个Channel保存住就行。当需要返回响应数据时就主动向持有的Channel写数据。具体请参照后面第4节。
关于客户端也是同理,在启动客户端之后要拿到Channel,当要主动发送数据时就向Channel中写入。
EventLoopGroup group = new NioEventLoopGroup();
Bootstrap b = new Bootstrap();
b.group(group)
.channel(NioSocketChannel.class)
.remoteAddress(host, port)
.handler(new ChannelInitializer<SocketChannel>() {
@Override
protected void initChannel(SocketChannel ch) throws Exception {
ch.pipeline().addLast(...);
}
});
try {
ChannelFuture future = b.connect().sync();
this.channel = future.channel();
}
catch (InterruptedException e) {
throw new IllegalStateException("Error when start netty client: addr=[" + addr + "]", e);
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
1.2 减少线程数
线程太多的话每个线程得到的时间片就少,CPU要让各个线程都有机会执行就要切换,切换就要不断保存和还原线程的上下文现场。于是检查Netty的I/O worker的EventLoopGroup。之前在《Netty 4源码解析:服务端启动》中曾经分析过,EventLoopGroup默认的线程数是CPU核数的二倍。所以手动配置NioEventLoopGroup的线程数,减少一些I/O线程。
private void doStartNettyServer(int port) throws InterruptedException {
EventLoopGroup bossGroup = new NioEventLoopGroup();
EventLoopGroup workerGroup = new NioEventLoopGroup(4);
try {
ServerBootstrap b = new ServerBootstrap()
.group(bossGroup, workerGroup)
.channel(NioServerSocketChannel.class)
.localAddress(port)
.childHandler(new ChannelInitializer<SocketChannel>() {
@Override
public void initChannel(SocketChannel ch) throws Exception {
ch.pipeline().addLast(...);
}
});
// Bind and start to accept incoming connections.
ChannelFuture f = b.bind(port).sync();
// Wait until the server socket is closed.
f.channel().closeFuture().sync();
} finally {
bossGroup.shutdownGracefully();
workerGroup.shutdownGracefully();
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
此外因为还用了Akka作为业务线程池,所以还看了下如何修改Akka的默认配置。方法是新建一个叫做application.conf的配置文件,我们创建ActorSystem时会自动加载这个配置文件,下面的配置文件中定制了一个dispatcher:
my-dispatcher {
# Dispatcher is the name of the event-based dispatcher
type = Dispatcher
mailbox-type = "akka.dispatch.SingleConsumerOnlyUnboundedMailbox"
# What kind of ExecutionService to use
executor = "fork-join-executor"
# Configuration for the fork join pool
fork-join-executor {
# Min number of threads to cap factor-based parallelism number to
parallelism-min = 2
# Parallelism (threads) ... ceil(available processors * factor)
parallelism-factor = 1.0
# Max number of threads to cap factor-based parallelism number to
parallelism-max = 16
}
# Throughput defines the maximum number of messages to be
# processed per actor before the thread jumps to the next actor.
# Set to 1 for as fair as possible.
throughput = 100
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
简单来说,最关键的几个配置项是:
- parallelism-factor:决定线程池的大小(竟然不是parallelism-max)。
- throughput:决定coroutine的切换频率,1是最为频繁也最为公平的设置。
因为本篇主要是介绍Netty的,所以具体含义就详细介绍了,请参考官方文档中对Dispatcher和Mailbox的介绍。创建特定Dispatcher的Akka很简单,以下是创建类型化Actor时指定Dispatcher的方法。
TypedActor.get(system).typedActorOf(
new TypedProps<MyActorImpl>(
MyActor.class,
new Creator<MyActorImpl>() {
@Override
public MyActorImpl create() throws Exception {
return new MyActorImpl(XXX);
}
}
).withDispatcher("my-dispatcher")
);
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
1.3 去业务线程池
尽管上面做了种种改进配置,用jstack查看线程配置确实生效了,但Context Switch的状况并没有好转。于是干脆去掉Akka实现的业务线程池,彻底减少线程上下文的切换。发现CS从30w+一下子降到了16w!费了好大力气在万能的StackOverflow上查到了一篇文章,其中一句话点醒了我:
And if the recommendation is not to block in the event loop, then this can be done in an application thread. But that would imply an extra context switch. This extra context switch may not be acceptable to latency sensitive applaications.
有了线索就赶紧去查Netty源码,发现的确像调用channel.write()操作不是在当前线程上执行。Netty内部统一使用executor.inEventLoop()判断当前线程是否是EventLoopGroup的线程,否则会包装好Task交给内部线程池执行:
private void write(Object msg, boolean flush, ChannelPromise promise) {
AbstractChannelHandlerContext next = findContextOutbound();
EventExecutor executor = next.executor();
if (executor.inEventLoop()) {
next.invokeWrite(msg, promise);
if (flush) {
next.invokeFlush();
}
} else {
int size = channel.estimatorHandle().size(msg);
if (size > 0) {
ChannelOutboundBuffer buffer = channel.unsafe().outboundBuffer();
// Check for null as it may be set to null if the channel is closed already
if (buffer != null) {
buffer.incrementPendingOutboundBytes(size);
}
}
Runnable task;
if (flush) {
task = WriteAndFlushTask.newInstance(next, msg, size, promise);
} else {
task = WriteTask.newInstance(next, msg, size, promise);
}
safeExecute(executor, task, promise, msg);
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
业务线程池原来是把双刃剑。虽然将任务交给业务线程池异步执行降低了Netty的I/O线程的占用时间、减轻了压力,但同时业务线程池增加了线程上下文切换的次数。通过上述这些优化手段,终于将压测时的CS从每秒30w+降到了8w左右,效果还是挺明显的!
2.系统调用开销
系统调用一般会涉及到从User Space到Kernel Space的模态转换(Mode Transition或Mode Switch)。这种转换也是有一定开销的。
Mode Switch vs. Context Switch
StackOverflow上果然什么问题都有。前面介绍过了线程的上下文切换,那它与内核态和用户态的切换是什么关系?模态切换算是CS的一种吗?Does there have to be a mode switch for something to qualify as a context switch?回答了这个问题:
“A mode switch happens inside one process. A context switch involves more than one process (or thread). Context switch happens only in kernel mode. If context switching happens between two user mode processes, first cpu has to change to kernel mode, perform context switch, return back to user mode and so on. So there has to be a mode switch associated with a context switch. But a context switch doesn’t imply a mode switch (could be done by the hardware alone). A mode switch does not require a context switch either.”
Context Switch必须在内核中完成,原理简单说就是主动触发一个软中断(类似被动被硬件触发的硬中断),所以一般Context Switch都会伴随Mode Switch。然而有些硬件也可以直接完成(不是很懂了),有些CPU甚至没有我们常说Ring 0 ~ 3的特权级概念。而Mode Switch则与Context Switch更是无关了,按照Wiki上的说法硬要扯上关系的话也只能说有的系统里可能在Mode Switch中发生Context Switch。
Netty涉及的系统调用最多的就是网络通信操作了,所以为了降低系统调用的频度,最直接的方法就是缓冲输出内容,达到一定的数据大小、写入次数或时间间隔时才flush缓冲区。
对于缓冲区大小不足,写入速度过快等问题,Netty提供了writeBufferLowWaterMark和writeBufferHighWaterMark选项,当缓冲区达到一定大小时则不能写入,避免被撑爆。感觉跟Netty提供的Traffic Shaping流量整形功能有点像呢。具体还未深入研究,感兴趣的同学可以自行学习一下。
3.Zero Copy实现
《Netty权威指南(第二版)》中专门有一节介绍Netty的Zero Copy,但针对的是Netty内部的零拷贝功能。我们这里想谈的是如何在应用代码中实现Zero Copy,最典型的应用场景就是消息透传。因为透传不需要完整解析消息,只需要知道消息要转发给下游哪个系统就足够了。所以透传时,我们可以只解析出部分消息,消息整体还原封不动地放在Direct Buffer里,最后直接将它写入到连接下游系统的Channel中。所以应用层的Zero Copy实现就分为两部分:Direct Buffer配置和Buffer的零拷贝传递。
3.1 内存池
使用Netty带来的又一个好处就是内存管理。只需一行简单的配置,就能获得到内存池带来的好处。在底层,Netty实现了一个Java版的Jemalloc内存管理库(还记得Redis自带的那个吗),为我们做完了所有“脏活累活”!
ServerBootstrap b = new ServerBootstrap()
.group(bossGroup, workerGroup)
.channel(NioServerSocketChannel.class)
.localAddress(port)
.childOption(ChannelOption.ALLOCATOR, PooledByteBufAllocator.DEFAULT)
.childHandler(new ChannelInitializer<SocketChannel>() {
@Override
public void initChannel(SocketChannel ch) throws Exception {
ch.pipeline().addLast(...);
}
});
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
3.2 应用层的Zero Copy
默认情况下,Netty会自动释放ByteBuf。也就是说当我们覆写的channelRead0()返回时,ByteBuf就结束了它的使命,被Netty自动释放掉(如果是池化的就可会被放回到内存池中)。
public abstract class SimpleChannelInboundHandler<I> extends ChannelInboundHandlerAdapter {
@Override
public void channelRead(ChannelHandlerContext ctx, Object msg) throws Exception {
boolean release = true;
try {
if (acceptInboundMessage(msg)) {
@SuppressWarnings("unchecked")
I imsg = (I) msg;
channelRead0(ctx, imsg);
} else {
release = false;
ctx.fireChannelRead(msg);
}
} finally {
if (autoRelease && release) {
ReferenceCountUtil.release(msg);
}
}
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
因为Netty是用引用计数的方式来判断是否回收的,所以要想继续使用ByteBuf而不让Netty释放的话,就要增加它的引用计数。只要我们在ChannelPipeline中的任意一个Handler中调用ByteBuf.retain()将引用计数加1,Netty就不会释放掉它了。我们在连接下游的客户端的Encoder中发送消息成功后再释放掉,这样就达到了零拷贝透传的效果:
public class RespEncoder extends MessageToByteEncoder<Resp> {
@Override
protected void encode(ChannelHandlerContext ctx, Msg msg, ByteBuf out) throws Exception {
// Raw in Msg is retained ByteBuf
out.writeBytes(msg.getRaw(), 0, msg.getRaw().readerIndex());
msg.getRaw().release();
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
4.并发下的状态处理
前面第1.1节介绍的异步写入持有的Channel和第2节介绍的根据一定规则flush缓冲区等等,都涉及到状态的保存。如果要并发访问这些状态的话,就要提防并发的race condition问题,避免更新冲突、丢失等等。
4.1 Channel保存
在Netty服务端的Handler里如何持有Channel呢?我是这样做的,在channelActive()或第一次进入channelRead0()时创建一个Session对象持有Channel。因为之前在《Netty 4源码解析:请求处理》中曾经分析过Netty 4的线程模型:多个客户端可能会对应一个EventLoop线程,但对于一个客户端来说只能对应一个EventLoop线程。每个客户端都对应自己的Handler实例,并且一直使用到连接断开。
public class FrontendHandler extends SimpleChannelInboundHandler<Msg> {
private Session session;
@Override
public void channelActive(ChannelHandlerContext ctx) throws Exception {
session = factory.createSession(ctx.channel());
super.channelActive(ctx);
}
@Override
protected void channelRead0(final ChannelHandlerContext ctx, Msg msg) throws Exception {
session.handleRequest(msg);
}
@Override
public void channelInactive(ChannelHandlerContext ctx) throws Exception {
session = null;
super.channelInactive(ctx);
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
4.2 Decoder状态
因为网络粘包拆包等因素,Decoder不可避免的要保存一些解析过程的中间状态。因为Netty对于每个客户端的生命周期内会一直使用同一个Decoder实例,所以解析完成后一定要重置中间状态,避免后续解析错误。
public class RespDecoder extends ReplayingDecoder {
public MsgDecoder() {
doCleanUp();
}
@Override
protected void decode(ChannelHandlerContext ctx, ByteBuf in, List<Object> out)
throws Exception {
if (doParseMsg(in)) {
doSendToHandler(out);
doCleanUp();
}
}
}
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
5.总结
5.1 多变的Netty
总结之前先吐槽一下,令人又爱又恨的Netty更新速度。从Netty 3到Netty 4,API发生了一次“大地震”,好多网上的示例程序都是基于Netty 3,所以学习Netty 4时发现好多例子都跑不起来了。除了API,Netty内部的线程模型等等变化就更不用说了。本以为用上了Netty 4就可以安心了,结果Netty 5的线程模型又-变-了!看看官方文档里的说法吧,升级的话又要注意了。
Even more flexible thread model
In Netty 4.x each EventLoop is tightly coupled with a fixed thread that executes all I/O events of its registered Channels and any tasks submitted to it. Starting with version 5.0 an EventLoop does no longer use threads directly but instead makes use of an Executor abstraction. That is, it takes an Executor object as a parameter in its constructor and instead of polling for I/O events in an endless loop each iteration is now a task that is submitted to this Executor. Netty 4.x would simply spawn its own threads and completely ignore the fact that it’s part of a larger system. Starting with Netty 5.0, developers can run Netty and the rest of the system in the same thread pool and potentially improve performance by applying better scheduling strategies and through less scheduling overhead (due to fewer threads). It shall be mentioned, that this change does not in any way affect the way ChannelHandlers are developed. From a developer’s point of view, the only thing that changes is that it’s no longer guaranteed that a ChannelHandler will always be executed by the same thread. It is, however, guaranteed that it will never be executed by two or more threads at the same time. Furthermore, Netty will also take care of any memory visibility issues that might occur. So there’s no need to worry about thread-safety and volatile variables within a ChannelHandler.
根据官方文档的说法,Netty不再保证特定的Handler实例在运行时一定对应一个线程,所以,在Handler中用ThreadLocal的话就是比较危险的写法了!
5.2 高并发编程技巧
经过上面的种种琢磨和努力,tps终于从几千达到了5w左右,学到了很多之前不懂的网络编程和性能优化的知识,还是很有成就感的!总结一下,高并发中间件的优化策略有:
- 线程数控制:高并发下如果线程较多时,Context Switch会非常明显,超过CPU核心数的线程不会带来任何好处。不是特别耗时的操作的话,业务线程池也是有害无益的。Netty 5为我们提供了指定底层线程池的机会,这样能更好的控制整个中间件的线程数和调度策略。
- 非阻塞I/O操作:要想线程少还多做事,避免阻塞是一定要做的。
- 减少系统调用:虽然Mode Switch比Context Switch的开销要小得多,但我们还是要尽量减少频繁的syscall。
- 数据零拷贝:从内核空间的Direct Buffer拷贝到用户空间,每次透传都拷贝的话累积起来是个不小的开销。
- 共享状态保护:中间件内部的并发处理也是决定性能的关键。