100万并发连接服务器笔记之Java Netty处理1M连接会怎么样

转载:http://www.blogjava.net/yongboy/archive/2013/05/13/399203.html

前言

每一种该语言在某些极限情况下的表现一般都不太一样，那么我常用的Java语言，在达到100万个并发连接情况下，会怎么样呢，有些好奇，更有些期盼。
这次使用经常使用的顺手的netty NIO框架（netty-3.6.5.Final），封装的很好，接口很全面，就像它现在的域名 netty.io，专注于网络IO。
整个过程没有什么技术含量，浅显分析过就更显得有些枯燥无聊，准备好，硬着头皮吧。

测试服务器配置

运行在VMWare Workstation 9中，64位Centos 6.2系统，分配14.9G内存左右，4核。
已安装有Java7版本：

java version "1.7.0_21"
Java(TM) SE Runtime Environment (build 1.7.0_21-b11)
Java HotSpot(TM) 64-Bit Server VM (build 23.21-b01, mixed mode)

测试端

测试端和以前一样的程序，翻看前几篇博客就可以看到client5.c的源码。

在/etc/sysctl.conf中添加如下配置：

fs.file-max = 1048576
net.ipv4.ip_local_port_range = 1024 65535
net.ipv4.tcp_mem = 786432 2097152 3145728
net.ipv4.tcp_rmem = 4096 4096 16777216
net.ipv4.tcp_wmem = 4096 4096 16777216

net.ipv4.tcp_tw_reuse = 1
net.ipv4.tcp_tw_recycle = 1

服务器程序

这次也是很简单呐，没有业务功能，客户端HTTP请求，服务端输出chunked编码内容。

入口HttpChunkedServer.java：

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465

package com . test . server ;   import static org . jboss . netty . channel . Channels . pipeline ;   import java.net.InetSocketAddress ; import java.util.concurrent.Executors ;   import org.jboss.netty.bootstrap.ServerBootstrap ; import org.jboss.netty.channel.ChannelPipeline ; import org.jboss.netty.channel.ChannelPipelineFactory ; import org.jboss.netty.channel.socket.nio.NioServerSocketChannelFactory ; import org.jboss.netty.handler.codec.http.HttpChunkAggregator ; import org.jboss.netty.handler.codec.http.HttpRequestDecoder ; import org.jboss.netty.handler.codec.http.HttpResponseEncoder ; import org.jboss.netty.handler.stream.ChunkedWriteHandler ;   public class HttpChunkedServer { private final int port ;   public HttpChunkedServer ( int port ) { this . port = port ; }   public void run () { // Configure the server. ServerBootstrap bootstrap = new ServerBootstrap ( new NioServerSocketChannelFactory ( Executors . newCachedThreadPool (), Executors . newCachedThreadPool ()));   // Set up the event pipeline factory. bootstrap . setPipelineFactory ( new ChannelPipelineFactory () { public ChannelPipeline getPipeline () throws Exception { ChannelPipeline pipeline = pipeline ();   pipeline . addLast ( "decoder" , new HttpRequestDecoder ()); pipeline . addLast ( "aggregator" , new HttpChunkAggregator ( 65536 )); pipeline . addLast ( "encoder" , new HttpResponseEncoder ()); pipeline . addLast ( "chunkedWriter" , new ChunkedWriteHandler ());   pipeline . addLast ( "handler" , new HttpChunkedServerHandler ()); return pipeline ; } });   bootstrap . setOption ( "child.reuseAddress" , true ); bootstrap . setOption ( "child.tcpNoDelay" , true ); bootstrap . setOption ( "child.keepAlive" , true );   // Bind and start to accept incoming connections. bootstrap . bind ( new InetSocketAddress ( port )); }   public static void main ( String [] args ) { int port ; if ( args . length > 0 ) { port = Integer . parseInt ( args [ 0 ]); } else { port = 8080 ; }   System . out . format ( "server start with port %d \n" , port ); new HttpChunkedServer ( port ). run (); } }

view raw HttpChunkedServer.java This Gist brought to you by  GitHub.

唯一的自定义处理器HttpChunkedServerHandler.java:

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120

package com . test . server ;   import static org . jboss . netty . handler . codec . http . HttpHeaders . Names . CONTENT_TYPE ; import static org . jboss . netty . handler . codec . http . HttpMethod . GET ; import static org . jboss . netty . handler . codec . http . HttpResponseStatus . BAD_REQUEST ; import static org . jboss . netty . handler . codec . http . HttpResponseStatus . METHOD_NOT_ALLOWED ; import static org . jboss . netty . handler . codec . http . HttpResponseStatus . OK ; import static org . jboss . netty . handler . codec . http . HttpVersion . HTTP_1_1 ;   import java.util.concurrent.atomic.AtomicInteger ;   import org.jboss.netty.buffer.ChannelBuffer ; import org.jboss.netty.buffer.ChannelBuffers ; import org.jboss.netty.channel.Channel ; import org.jboss.netty.channel.ChannelFutureListener ; import org.jboss.netty.channel.ChannelHandlerContext ; import org.jboss.netty.channel.ChannelStateEvent ; import org.jboss.netty.channel.ExceptionEvent ; import org.jboss.netty.channel.MessageEvent ; import org.jboss.netty.channel.SimpleChannelUpstreamHandler ; import org.jboss.netty.handler.codec.frame.TooLongFrameException ; import org.jboss.netty.handler.codec.http.DefaultHttpChunk ; import org.jboss.netty.handler.codec.http.DefaultHttpResponse ; import org.jboss.netty.handler.codec.http.HttpChunk ; import org.jboss.netty.handler.codec.http.HttpHeaders ; import org.jboss.netty.handler.codec.http.HttpRequest ; import org.jboss.netty.handler.codec.http.HttpResponse ; import org.jboss.netty.handler.codec.http.HttpResponseStatus ; import org.jboss.netty.util.CharsetUtil ;   public class HttpChunkedServerHandler extends SimpleChannelUpstreamHandler { private static final AtomicInteger count = new AtomicInteger ( 0 );   private void increment () { System . out . format ( "online user %d\n" , count . incrementAndGet ()); }   private void decrement () { if ( count . get () <= 0 ) { System . out . format ( "~online user %d\n" , 0 ); } else { System . out . format ( "~online user %d\n" , count . decrementAndGet ()); } }   @Override public void messageReceived ( ChannelHandlerContext ctx , MessageEvent e ) throws Exception { HttpRequest request = ( HttpRequest ) e . getMessage (); if ( request . getMethod () != GET ) { sendError ( ctx , METHOD_NOT_ALLOWED ); return ; }   sendPrepare ( ctx ); increment (); }   @Override public void channelDisconnected ( ChannelHandlerContext ctx , ChannelStateEvent e ) throws Exception { decrement (); super . channelDisconnected ( ctx , e ); }   @Override public void exceptionCaught ( ChannelHandlerContext ctx , ExceptionEvent e ) throws Exception { Throwable cause = e . getCause (); if ( cause instanceof TooLongFrameException ) { sendError ( ctx , BAD_REQUEST ); return ; } }   private static void sendError ( ChannelHandlerContext ctx , HttpResponseStatus status ) { HttpResponse response = new DefaultHttpResponse ( HTTP_1_1 , status ); response . setHeader ( CONTENT_TYPE , "text/plain; charset=UTF-8" ); response . setContent ( ChannelBuffers . copiedBuffer ( "Failure: " + status . toString () + "\r\n" , CharsetUtil . UTF_8 ));   // Close the connection as soon as the error message is sent. ctx . getChannel (). write ( response ) . addListener ( ChannelFutureListener . CLOSE ); }   private void sendPrepare ( ChannelHandlerContext ctx ) { HttpResponse response = new DefaultHttpResponse ( HTTP_1_1 , OK ); response . setChunked ( true ); response . setHeader ( HttpHeaders . Names . CONTENT_TYPE , "text/html; charset=UTF-8" ); response . addHeader ( HttpHeaders . Names . CONNECTION , HttpHeaders . Values . KEEP_ALIVE ); response . setHeader ( HttpHeaders . Names . TRANSFER_ENCODING , HttpHeaders . Values . CHUNKED );   Channel chan = ctx . getChannel (); chan . write ( response );   // 缓冲必须凑够256字节，浏览器端才能够正常接收 ... StringBuilder builder = new StringBuilder (); builder . append ( "<html><body><script>var _ = function (msg) { parent.s._(msg, document); };</script>" ); int leftChars = 256 - builder . length (); for ( int i = 0 ; i < leftChars ; i ++) { builder . append ( " " ); }   writeStringChunk ( chan , builder . toString ()); }   private void writeStringChunk ( Channel channel , String data ) { ChannelBuffer chunkContent = ChannelBuffers . dynamicBuffer ( channel . getConfig (). getBufferFactory ()); chunkContent . writeBytes ( data . getBytes ()); HttpChunk chunk = new DefaultHttpChunk ( chunkContent );   channel . write ( chunk ); } }

view raw HttpChunkedServerHandler.java This Gist brought to you by  GitHub.

启动脚本start.sh

12

set CLASSPATH =. nohup java -server -Xmx6G -Xms6G -Xmn600M -XX:PermSize =50M -XX:MaxPermSize =50M -Xss256K -XX:+DisableExplicitGC -XX:SurvivorRatio =1 -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:+CMSParallelRemarkEnabled -XX:+UseCMSCompactAtFullCollection -XX:CMSFullGCsBeforeCompaction =0 -XX:+CMSClassUnloadingEnabled -XX:LargePageSizeInBytes =128M -XX:+UseFastAccessorMethods -XX:+UseCMSInitiatingOccupancyOnly -XX:CMSInitiatingOccupancyFraction =80 -XX:SoftRefLRUPolicyMSPerMB =0 -XX:+PrintClassHistogram -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintHeapAtGC -Xloggc:gc.log -Djava.ext.dirs =lib com.test.server.HttpChunkedServer 8000 >server.out 2>&1 &

view raw start.sh This Gist brought to you by  GitHub.

达到100万并发连接时的一些信息

每次服务器端达到一百万个并发持久连接之后，然后关掉测试端程序，断开所有的连接，等到服务器端日志输出在线用户为0时，再次重复以上步骤。在这反反复复的情况下，观察内存等信息的一些情况。以某次断开所有测试端为例后，当前系统占用为（设置为list_free_1）:

                  total       used       free     shared    buffers     cached
     Mem:         15189       7736       7453          0         18        120
     -/+ buffers/cache:       7597       7592
     Swap:         4095        948       3147

通过top观察，其进程相关信息

    PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                                                       
   4925 root      20   0 8206m 4.3g 2776 S  0.3 28.8  50:18.66 java

在启动脚本start.sh中，我们设置堆内存为6G。

ps aux|grep java命令获得信息：

  root      4925 38.0 28.8 8403444 4484764 ?     Sl   15:26  50:18 java -server...HttpChunkedServer 8000

RSS占用内存为4484764K/1024K=4379M

然后再次启动测试端，在服务器接收到online user 1023749时，ps aux|grep java内容为：

  root      4925 43.6 28.4 8403444 4422824 ?     Sl   15:26  62:53 java -server...

查看当前网络信息统计

  ss -s
  Total: 1024050 (kernel 1024084)
  TCP:   1023769 (estab 1023754, closed 2, orphaned 0, synrecv 0, timewait 0/0), ports 12

  Transport Total     IP        IPv6
  *    1024084   -         -        
  RAW     0         0         0        
  UDP     7         6         1        
  TCP     1023767   12        1023755  
  INET    1023774   18        1023756  
  FRAG    0         0         0    

通过top查看一下

  top -p 4925
  top - 17:51:30 up  3:02,  4 users,  load average: 1.03, 1.80, 1.19
  Tasks:   1 total,   0 running,   1 sleeping,   0 stopped,   0 zombie
  Cpu0  :  0.9%us,  2.6%sy,  0.0%ni, 52.9%id,  1.0%wa, 13.6%hi, 29.0%si,  0.0%st
  Cpu1  :  1.4%us,  4.5%sy,  0.0%ni, 80.1%id,  1.9%wa,  0.0%hi, 12.0%si,  0.0%st
  Cpu2  :  1.5%us,  4.4%sy,  0.0%ni, 80.5%id,  4.3%wa,  0.0%hi,  9.3%si,  0.0%st
  Cpu3  :  1.9%us,  4.4%sy,  0.0%ni, 84.4%id,  3.2%wa,  0.0%hi,  6.2%si,  0.0%st
  Mem:  15554336k total, 15268728k used,   285608k free,     3904k buffers
  Swap:  4194296k total,  1082592k used,  3111704k free,    37968k cached

    PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                                                       
   4925 root      20   0 8206m 4.2g 2220 S  3.3 28.4  62:53.66 java

四核都被占用了，每一个核心不太平均。这是在虚拟机中得到结果，可能真实服务器会更好一些。 因为不是CPU密集型应用，CPU不是问题，无须多加关注。

系统内存状况

  free -m
               total       used       free     shared    buffers     cached
  Mem:         15189      14926        263          0          5         56
  -/+ buffers/cache:      14864        324
  Swap:         4095       1057       3038

物理内存已经无法满足要求了，占用了1057M虚拟内存。

查看一下堆内存情况

  jmap -heap 4925
  Attaching to process ID 4925, please wait...
  Debugger attached successfully.
  Server compiler detected.
  JVM version is 23.21-b01

  using parallel threads in the new generation.
  using thread-local object allocation.
  Concurrent Mark-Sweep GC

  Heap Configuration:
     MinHeapFreeRatio = 40
     MaxHeapFreeRatio = 70
     MaxHeapSize      = 6442450944 (6144.0MB)
     NewSize          = 629145600 (600.0MB)
     MaxNewSize       = 629145600 (600.0MB)
     OldSize          = 5439488 (5.1875MB)
     NewRatio         = 2
     SurvivorRatio    = 1
     PermSize         = 52428800 (50.0MB)
     MaxPermSize      = 52428800 (50.0MB)
     G1HeapRegionSize = 0 (0.0MB)

  Heap Usage:
  New Generation (Eden + 1 Survivor Space):
     capacity = 419430400 (400.0MB)
     used     = 308798864 (294.49354553222656MB)
     free     = 110631536 (105.50645446777344MB)
     73.62338638305664% used
  Eden Space:
     capacity = 209715200 (200.0MB)
     used     = 103375232 (98.5863037109375MB)
     free     = 106339968 (101.4136962890625MB)
     49.29315185546875% used
  From Space:
     capacity = 209715200 (200.0MB)
     used     = 205423632 (195.90724182128906MB)
     free     = 4291568 (4.0927581787109375MB)
     97.95362091064453% used
  To Space:
     capacity = 209715200 (200.0MB)
     used     = 0 (0.0MB)
     free     = 209715200 (200.0MB)
     0.0% used
  concurrent mark-sweep generation:
     capacity = 5813305344 (5544.0MB)
     used     = 4213515472 (4018.321487426758MB)
     free     = 1599789872 (1525.6785125732422MB)
     72.48054631000646% used
  Perm Generation:
     capacity = 52428800 (50.0MB)
     used     = 5505696 (5.250640869140625MB)
     free     = 46923104 (44.749359130859375MB)
     10.50128173828125% used

  1439 interned Strings occupying 110936 bytes.

老生代占用内存为72%，较为合理，毕竟系统已经处理100万个连接。

再次断开所有测试端，看看系统内存（free -m）

               total       used       free     shared    buffers     cached
  Mem:         15189       7723       7466          0         13        120
  -/+ buffers/cache:       7589       7599
  Swap:         4095        950       3145

记为list_free_2。

list_free_1和list_free_2两次都释放后的内存比较结果，系统可用物理已经内存已经降到7589M，先前可是7597M物理内存。
总之，我们的JAVA测试程序在内存占用方面已经，最低需要7589 + 950 = 8.6G内存为最低需求内存吧。

GC日志

我们在启动脚本处设置的一大串参数，到底是否达到目标，还得从gc日志处获得具体效果，推荐使用GCViewer。

GC事件概览：

其它:
  

总之：

• 只进行了一次Full GC，代价太高，停顿了12秒。
• PartNew成为了停顿大户，导致整个系统停顿了41秒之久，不可接受。
• 当前JVM调优喜忧参半，还得继续努力等

小结

Java与与Erlang、C相比，比较麻烦的事情，需要在程序一开始就得准备好它的堆栈到底需要多大空间，换个说法就是JVM启动参数设置堆内存大小，设置合适的垃圾回收机制，若以后程序需要更多内存，需停止程序，编辑启动参数，然后再次启动。总之一句话，就是麻烦。单单JVM的调优，就得持续不断的根据检测、信息、日志等进行适当微调。

• JVM需要提前指定堆大小，相比Erlang/C，这可能是个麻烦
• GC(垃圾回收)，相对比麻烦，需要持续不断的根据日志、JVM堆栈信息、运行时情况进行JVM参数微调
• 设置一个最大连接目标，多次测试达到顶峰，然后释放所有连接，反复观察内存占用，获得一个较为合适的系统运行内存值
• Eclipse Memory Analyzer结合jmap导出堆栈DUMP文件，分析内存泄漏，还是很方便的
• 想修改运行时内容，或者称之为热加载，默认不可能
• 真实机器上会有更好的反映

吐槽一下：
JAVA OSGI，相对比Erlang来说，需要人转换思路，不是那么原生的东西，总是有些别扭，社区或商业公司对此的修修补补，不过是实现一些面向对象所不具备的热加载的企业特性。

测试源代码，下载just_test。