Java I/O vs NIO: Quick Benchmark Comparison

Question

I recently read that in newer computers Java's I/O performs better than NIO because of the new availability of multi-core machines.

I ran a quick test comparing the transfer time of I/O and NIO over the LAN using the localhost loopback address.

Note: This is using JDK 7

The results (3 trials):

I/O transfers averaged 21789.3ms

NIO transfers averaged 22771.0ms

It is also worth noting that CPU usage appeared to be around 10% higher on each NIO transfer as compared to the I/O.

My question for you is if my comparison code is fair? Did I write good/equal I/O and NIO code? If not, how can I improve and re-run this test?

    public static void main(String[] args) {
    System.out.println("Initiating test sequence...");
    new Thread(new Client()).start();
    try {
        System.out.println("Server I/O initiating...");
        ServerSocket server = new ServerSocket(5555);
        Socket sock = server.accept();
        System.out.println("Server connected to client successfully");
        InputStream is = sock.getInputStream();
        File output = new File("C:/test_root/video.avi");
        FileOutputStream fos = new FileOutputStream(output);
        byte[] data = new byte[1024];
        int len=0;
        System.out.println("Server initiating transfer - Timer starting");
        long start = System.currentTimeMillis();
        while((len=is.read(data))>0) {
            fos.write(data, 0, len);
            fos.flush();
        }
        fos.close();
        is.close();
        sock.close();
        server.close();
        long end = System.currentTimeMillis();
        System.out.println("Network I/O transfer time = "+(end-start)+"ms");

        System.out.println("Server NIO initiating...");
        ServerSocketChannel serverChan = ServerSocketChannel.open();
        serverChan.bind(new InetSocketAddress(5555));
        SocketChannel chan = serverChan.accept();
        chan.configureBlocking(false);
        System.out.println("Server channel connected");
        FileChannel fc = (FileChannel) Files.newByteChannel(Paths.get("C:/test_root/video.avi"), StandardOpenOption.CREATE, StandardOpenOption.WRITE);
        ByteBuffer buff = ByteBuffer.allocate(1024);
        System.out.println("Server initiating transfer - Timer starting");
        start = System.currentTimeMillis();
        while(chan.read(buff)>=0 || buff.position() > 0) {
            buff.flip();
            fc.write(buff);
            buff.compact();
        }
        chan.close();
        fc.close();
        serverChan.close();
        end = System.currentTimeMillis();
        System.out.println("Network NIO transfer time = "+(end-start)+"ms");
    } catch (IOException e) {
        e.printStackTrace();
    }
    System.out.println("Test completed!");
}

static class Client implements Runnable {

    public void run() {
        try {
            System.out.println("Client I/O initiating...");
            Socket sock = new Socket("localhost", 5555);
            System.out.println("Client connected to server successfully!");
            OutputStream os = sock.getOutputStream();
            File input = new File(System.getProperty("user.home")+"/Documents/clip0025.avi");
            FileInputStream fis = new FileInputStream(input);
            byte[] data = new byte[1024];
            int len=0;
            int tot=0;
            int perc=0;
            while((len=fis.read(data))>0) {
                os.write(data, 0, len);
                os.flush();
                tot+=len;
                int prev = perc;
                perc = getPercentage(tot, input.length());
                if(perc !=prev && (perc == 10 || perc == 25 || perc == 50 || perc == 75 || perc == 98))
                    System.out.println("Client reporting: "+perc+"% read");
            }
            os.close();
            fis.close();
            sock.close();

            try {
                Thread.sleep(2000);
            } catch (InterruptedException e) {
                e.printStackTrace();
            }
            System.out.println("Client NIO initiating...");
            SocketChannel sc = SocketChannel.open();
            boolean connected = sc.connect(new InetSocketAddress("localhost",5555));
            if(!connected)
                connected = sc.finishConnect();
            if(!connected)
                throw(new IOException("Client failed to connect"));
            System.out.println("Client channel connected");
            sc.configureBlocking(false);
            FileChannel fc = (FileChannel) Files.newByteChannel(input.toPath(), StandardOpenOption.READ);
            ByteBuffer buff = ByteBuffer.allocate(1024);
            len=0;
            tot=0;
            while((len=fc.read(buff))>=0||buff.position()>0) {
                buff.flip();
                sc.write(buff);
                buff.compact();
                tot+=len;
                int prev = perc;
                perc = getPercentage(tot, input.length());
                if(perc !=prev && (perc == 10 || perc == 25 || perc == 50 || perc == 75 || perc == 98))
                    System.out.println("Client reporting: "+perc+"% read");
            }
            sc.close();
            fc.close();
        } catch (IOException e) {
            e.printStackTrace();
        }
    }
}

Additional Info:

Windows Vista (SP2) on Dell Studio XPS 435MT

1st gen i7 quad-core processor 2.67GHz

6GB RAM

64-bit architecture

Answer 1

Suggestions

• Try comparing blocking IO with blocking NIO. Your code will be shorter. If you are going to test IO, use IO on the client and server. If you are going to use NIO, use the same on both ends.
• Use direct ByteBuffers.
• Don't read/write to files as they are not part of the benchmark and can be much slower. Just pass around blank blocks of data.
• Try different block sizes, e.g. 8 KB.
• Consider the type of data you want to exchange. e.g. ByteBuffer may make reading int and long more efficient.
• Report numbers in terms of bandwidth. I would expect to see between 1-3 GB/sec over loop back.

http://vanillajava.blogspot.com/2010/07/java-nio-is-faster-than-java-io-for.html