Java网关服务-AIO（三）

前两节中，我们已经获取了body的总长度，剩下的就是读出body，处理请求

ChannelServerHandler

ChannelServerHandler即从channel中读取请求，也向channle输出结果，因此它实现了InboundHandler, OutboundHandler

/**
 * 读取请求的内容，业务处理
 */
public class ChannelServerHandler implements CompletionHandler<Integer, ByteBuffer>, InboundHandler, OutboundHandler {

    private final static Logger LOGGER = LoggerFactory.getLogger(ChannelServerHandler.class);

    private AsynchronousSocketChannel channel;

    public ChannelServerHandler(AsynchronousSocketChannel channel) {
        this.channel = channel;
    }

    public void completed(Integer result, ByteBuffer attachment) {
        //如果条件成立，说明客户端主动终止了TCP套接字，这时服务端终止就可以了
        if (result == -1) {
            System.out.println("remote is close");
            closeChannel();
            return;
        }

        Object resultData;
        String req = (String) read(channel, attachment);
        if (req == null) {
            closeChannel();
            return;
        }

        try {
            LOGGER.info("socket:{}", channel.getRemoteAddress());

            //同步处理请求
            RequestHandler requestHandler = ApplicationUtils.getBean(RequestHandler.class);
            resultData = requestHandler.execute(req);

        } catch (Throwable t) {
            resultData = Result.error("ERROR", Utils.error(t));
            LOGGER.error("调用接口失败", t);
        }

        if (resultData == null) {
            resultData = Result.failure("FAILURE", "调用失败,数据为空.");
        }
        try {
            String resultContent =  resultData instanceOf String ? (String) resultData : JSON.toJSONString(resultData);
            byte[] bytes = resultContent.getBytes("UTF-8");
            ByteBuffer writeBuffer = ByteBuffer.allocate(bytes.length);
            writeBuffer.put(bytes);
            writeBuffer.flip();

            write(channel, writeBuffer);
        } catch (Exception e) {
            LOGGER.error("对象转JSON失败,对象:{}", resultData, e);
        }

        closeChannel();
    }

    @Override
    public Object read(AsynchronousSocketChannel socketChannel, ByteBuffer in) {
        in.flip();
        byte[] body = new byte[in.remaining()];
        in.get(body);

        String req = null;
        try {
            req = new String(body, "UTF-8");
        } catch (UnsupportedEncodingException e) {
            e.printStackTrace();
        }
        return req;
    }

    @Override
    public Object write(AsynchronousSocketChannel socketChannel, ByteBuffer out) {
        //write，write操作结束后关闭通道
        channel.write(out, out, new CompletionHandler<Integer, ByteBuffer>() {
            @Override
            public void completed(Integer result, ByteBuffer attachment) {
                closeChannel();
            }

            @Override
            public void failed(Throwable exc, ByteBuffer attachment) {
                closeChannel();
            }
        });
        return null;
    }

    public void failed(Throwable exc, ByteBuffer attachment) {
        closeChannel();
    }

    private void closeChannel() {
        try {
            this.channel.close();
        } catch (IOException e) {
            e.printStackTrace();
        }
    }

}

读取body

        in.flip();
        byte[] body = new byte[in.remaining()];
        in.get(body);

        String req = null;
        try {
            req = new String(body, "UTF-8");
        } catch (UnsupportedEncodingException e) {
            e.printStackTrace();
        }
        return req;

in.remaining()

buffer中含有的字节数

客户端、服务端由于跨语言和经验问题，没有使用复杂的跨语言序列化技术，双方约定使用UTF-8编码，通过将body转换为String，最终获得了客户端传递的字符串。

处理请求

经过自定义的请求处理逻辑，同步处理，最终将响应编码后，发送给客户端，write操作结束后，关闭连接

总结

使用AIO开发服务端时，主要涉及

• 配置I/O事件完成的回调线程池
• 从accept -> read 到 向client端响应 write -> close,尽量使用CompletionHanlder来异步处理，不要在处理某个事件完成的线程中，同步的调用，如future.get()
• 如果是短连接，则需在write操作时注册write结束后的handler，在handler中关闭连接

扩展

长连接该如何处理

• 长连接意味着client可以发多次请求，由于多次请求被server执行的顺序是不可控的，可能后发的请求先响应，因此需要在请求和响应时，加requestId，据此对应到请求的结果
• 长连接不需要在write后关闭连接
• 长连接需要开发定时的ping-pong心跳消息
• 长连接在响应时比现在更复杂，也需要一个和请求类似或相同的协议来标识body长度

测试

测试用例

    /**
     * mvn -Dtest=com.jd.jshop.web.sdk.test.ClientTest#pingReqSocket test
     *
     * @throws IOException
     */
    @Test
    @PerfTest(invocations = 20000, threads = 50)
    public void pingReqSocket() throws IOException {

        byte[] content = "ping".getBytes("UTF-8");
        String result = sendReq(content);

        //断言 是否和预期一致
        Assert.assertEquals("pong", result);
    }

    private String sendReq(byte[] content) throws IOException {
        ByteBuffer writeBuffer = ByteBuffer.allocate(4 + content.length);
        writeBuffer.putInt(content.length);
        writeBuffer.put(content);
        writeBuffer.flip();

        Socket socket = new Socket();
        socket.connect(new InetSocketAddress("127.0.0.1", 9801));
        socket.getOutputStream().write(writeBuffer.array());
        socket.getOutputStream().flush();
        byte[] buf = new byte[1024];
        int len = 0;
        String result = null;
        while ((len = socket.getInputStream().read(buf)) != -1) {
            result = new String(buf, 0, len);
            System.out.println(result);
        }
        return result;
    }

测试的方法是，在服务器上建立socket连接，向server发送ping，server返回pong

测试服务器：centos, 2个物理核，4个逻辑核，内存16G

分析aio的实现：

在ping-pong测试中性能极高，优于并甩开netty

以下是使用Netty开发的server端的测试用例，可以和上面的图片对比一下

Measured invocations:    10,000
Thread Count:    10  

    Measured
(system)    Required
Execution time:    1,646 ms
Throughput:    6,075 / s
Min. latency:    0 ms
Average latency:    1 ms
Median:    2 ms
90%:    2 ms
Max latency:    26 ms   

============================

Started at:    Oct 16, 2018 5:27:03 PM
Measured invocations:    20,000
Thread Count:    20  

    Measured
(system)    Required
Execution time:    3,293 ms
Throughput:    6,073 / s
Min. latency:    0 ms
Average latency:    3 ms
Median:    3 ms
90%:    5 ms
Max latency:    54 ms   

============================

Started at:    Oct 16, 2018 5:28:24 PM
Measured invocations:    20,000
Thread Count:    10  

    Measured
(system)    Required
Execution time:    3,051 ms
Throughput:    6,555 / s
Min. latency:    0 ms
Average latency:    1 ms
Median:    1 ms
90%:    2 ms
Max latency:    44 ms   

============================

Started at:    Oct 16, 2018 5:30:06 PM
Measured invocations:    20,000
Thread Count:    50  

    Measured
(system)    Required
Execution time:    3,167 ms
Throughput:    6,315 / s
Min. latency:    0 ms
Average latency:    7 ms
Median:    7 ms
90%:    10 ms
Max latency:    64 ms

分析基于Netty的实现：

吞吐量：6000+/s

10个线程时：90%低于2ms,平均1ms

20个线程时：90%低于5ms，平均3ms

50个线程时：90%低于10ms，平均7ms

线程越多，性能越低

当前测试用例不太依赖内存

执行10000+次请求，建立10000+连接，要求服务器对单个进程fd限制打开,防止报too many open files导致测试用例执行失败

    ulimit -n 20240