"通过对等方重置连接"和"太多打开文件"异常将尝试扩展Vert.x服务器

Joe*_*ard 5 macos scalability websocket netty vert.x

我试图在Mac OS X上测试Java Vert.x网络服务器(基于Netty)的可扩展性.为此,我将连接限制增加到100万:

sudo sysctl -w kern.maxfiles=1000200
sudo sysctl -w kern.maxfilesperproc=1000100
sudo ulimit -n 1000000
Run Code Online (Sandbox Code Playgroud)

然后我创建了localhost别名:

for i in `seq 200 250`; do sudo ifconfig lo0 alias 172.16.123.$i ; done
Run Code Online (Sandbox Code Playgroud)

我现在能够连接到这些IP地址; 但是,在大约1000-2000个已建立的连接之后,我得到"通过对等连接重置"连接:

java.net.SocketException: Connection reset by peer
    at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
    at sun.nio.ch.SocketChannelImpl.finishConnect(SocketChannelImpl.java:712)
    at io.netty.channel.socket.nio.NioSocketChannel.doFinishConnect(NioSocketChannel.java:191)
    at io.netty.channel.nio.AbstractNioChannel$AbstractNioUnsafe.finishConnect(AbstractNioChannel.java:228)
    at io.netty.channel.nio.NioEventLoop.processSelectedKey(NioEventLoop.java:497)
    at io.netty.channel.nio.NioEventLoop.processSelectedKeysOptimized(NioEventLoop.java:447)
Run Code Online (Sandbox Code Playgroud)

它们之后是"太多打开文件"例外:

Caused by: java.net.SocketException: Too many open files
    at sun.nio.ch.Net.socket0(Native Method)
    at sun.nio.ch.Net.socket(Net.java:393)
    at sun.nio.ch.Net.socket(Net.java:386)
    at sun.nio.ch.SocketChannelImpl.<init>(SocketChannelImpl.java:104)
    at sun.nio.ch.SelectorProviderImpl.openSocketChannel(SelectorProviderImpl.java:60)
    at java.nio.channels.SocketChannel.open(SocketChannel.java:145)
    at io.netty.channel.socket.nio.NioSocketChannel.newSocket(NioSocketChannel.java:48)
Run Code Online (Sandbox Code Playgroud)

有趣的是,当我列出所有与netstat的连接时,我发现FIN_WAIT_1状态中有很多连接:

tcp4       0    213  172.16.123.230.55618   172.16.123.230.8877    FIN_WAIT_1 
tcp4       0    213  172.16.123.229.55624   172.16.123.229.8877    FIN_WAIT_1 
tcp4       0    213  172.16.123.228.55617   172.16.123.228.8877    FIN_WAIT_1 
tcp4       0    213  172.16.123.227.55616   172.16.123.227.8877    FIN_WAIT_1 
tcp4       0    213  172.16.123.226.55612   172.16.123.226.8877    FIN_WAIT_1 
Run Code Online (Sandbox Code Playgroud)

其他结果:

$ netstat -a | wc -l:
14282
$ lsof | wc -l
6922
Run Code Online (Sandbox Code Playgroud)

服务器代码(简化):

HttpServer server = vertx.createHttpServer();
server.websocketHandler(ws -> {
    connections++;
    ws.exceptionHandler(t -> t.printStackTrace());
    ws.closeHandler(h -> {
        connections--;
    });

    if (ws.path().equals("/app")) {
        ws.dataHandler(data -> {
            // ...
        });
    } else {
        ws.reject();
    }
}).listen(8877);
Run Code Online (Sandbox Code Playgroud)

客户代码(简化):

Handler<Void> connectHandler = new Handler<Void>() {
    @Override
    public void handle(Void e) {
        HttpClient client = vertx.createHttpClient().setHost(nextHost()).setPort(8877);

        client.connectWebsocket("/app", ws -> {
            connections++;
            ws.exceptionHandler(t -> t.printStackTrace());
            ws.closeHandler(h -> {
                connections--;
            });

            // ...
        });

        if (connections < 1000000) {
            vertx.runOnContext(this);
        }
    }
};
connectHandler.handle(null);
Run Code Online (Sandbox Code Playgroud)

两者都开始于(在同一台计算机上):

vertx run Class -cp classes:... -instances 8
Run Code Online (Sandbox Code Playgroud)

有没有人知道可能是什么问题?我还没有经历过这样的设置,也许我刚刚犯了一个简单的错误.它无助于增加接受积压.

更多的信息:

更新:

添加后

limit maxfiles 1000100 2000200
Run Code Online (Sandbox Code Playgroud)

到/etc/launchd.conf和

kern.maxfiles=2000400
kern.maxfilesperproc=1000200
Run Code Online (Sandbox Code Playgroud)

到/etc/sysctl.conf并重新启动计算机,它已经变得更好了.但是,在我获得"太多打开文件"之后,即使读取普通文件也是例外,尽管打开的文件不超过10000个:

Caused by: sun.nio.fs.UnixException: Too many open files
        at sun.nio.fs.UnixNativeDispatcher.getcwd(Native Method)
        at sun.nio.fs.UnixFileSystem.<init>(UnixFileSystem.java:67)
        at sun.nio.fs.BsdFileSystem.<init>(BsdFileSystem.java:41)
        at sun.nio.fs.MacOSXFileSystem.<init>(MacOSXFileSystem.java:44)
        at sun.nio.fs.MacOSXFileSystemProvider.newFileSystem(MacOSXFileSystemProvider.java:45)
        at sun.nio.fs.MacOSXFileSystemProvider.newFileSystem(MacOSXFileSystemProvider.java:38)
        at sun.nio.fs.UnixFileSystemProvider.<init>(UnixFileSystemProvider.java:56)
        at sun.nio.fs.BsdFileSystemProvider.<init>(BsdFileSystemProvider.java:36)
        at sun.nio.fs.MacOSXFileSystemProvider.<init>(MacOSXFileSystemProvider.java:40)
Run Code Online (Sandbox Code Playgroud)

更新2:

我也尝试了其他几个选项,但似乎有一个限制,正好是10180个连接,不能超过.除非有人有另一个想法,否则我可能只会切换到linux.

更新3:

添加JVM选项"-XX:-MaxFDLimit"之后,我现在可以打开16331个连接,直到出现以下异常:

java.net.BindException: Can't assign requested address
        at sun.nio.ch.Net.connect0(Native Method)
        at sun.nio.ch.Net.connect(Net.java:435)
        at sun.nio.ch.Net.connect(Net.java:427)
        at sun.nio.ch.SocketChannelImpl.connect(SocketChannelImpl.java:643)
        at io.netty.channel.socket.nio.NioSocketChannel.doConnect(NioSocketChannel.java:176)
Run Code Online (Sandbox Code Playgroud)

这可能是因为默认的端口范围(16383端口):

net.inet.ip.portrange.first: 49152
net.inet.ip.portrange.last: 65535
Run Code Online (Sandbox Code Playgroud)

但是,由于我使用了50个localhost alises,我原本希望能够打开16383*50 = 819150个连接.或者它是一个已知的限制,你不能使用相同的端口用于不同的别名?

更新4:

我试图用networksetup创建虚拟接口,但结果是一样的:

for i in `seq 200 250`; do sudo networksetup -createnetworkservice lo$i lo0 ; sudo networksetup -setmanual lo$i 172.16.123.$i 255.240.0.0 ; done
Run Code Online (Sandbox Code Playgroud)