The Netty library is a benchmark for high performance libraries. Netty is designed for scalability in particular, horizontal scalability is based on the Selector and it has highly tuned this facility.

This library is designed to be lower latency and support higher throughputs by employing techniques used in low latency trading systems.

Network current support TCP only.

Planned support for

• Shared Memory
• Unreliable UDP

The client sends a message to the server, the server immediately responds with the same message back to the client.

The full source code of this example can be found at:

net.openhft.performance.tests.network.SimpleServerAndClientTest.test

Below are some of the key parts of this code explained, in a bit more detail.

The TCPRegistry is most useful for unit tests, it allows you to either provide a true host and port, say "localhost:8080" or if you would rather let the application allocate you a free port at random, you can just provide a text reference to the port, such as, "host.port", you can provide any text you want. It will always be taken as a reference. That is unless its correctly formed like "hostname:port”, then it will use the exact host and port you provide. The reason we offer this functionality is quite often in unit tests you wish to start a test via loopback, followed often by another test, if the first test does not shut down correctly it can impact on the second test. Giving each test a unique port is one solution, but then managing those ports can become a problem in its self. So we created the TCPRegistry which manages those ports for you, when you come to clean up at the end of each test, all you have to do is call TCPRegistry.reset() and it will ensure that any remaining ports that are open will be closed.

// this the name of a reference to the host name and port,
// allocated automatically when to a free port on localhost
final String desc = "host.port";
TCPRegistry.createServerSocketChannelFor(desc);

// we use an event loop rather than lots of threads
EventLoop eg = new EventGroup(true);
eg.start();

The server is configured with TextWire, so the client must also be configured with TextWire. The port that we will use will be ( in this example ) determined by the TCP Registry, of course in a real life production environment you may decide not to use the TcpRegistry or if you still use the TcpRegistry you can use a fixed host:port.

final String expectedMessage = "<my message>";
AcceptorEventHandler eah = new AcceptorEventHandler(desc,
    () -> new WireEchoRequestHandler(WireType.TEXT), VanillaSessionDetails::new, 0, 0);
eg.addHandler(eah);
final SocketChannel sc = TCPRegistry.createSocketChannel(desc);
sc.configureBlocking(false);

The server code that processes a message, in this simple example we receive and update a message and then immediately send back a response, however there are other solutions that can be implemented using Chronicle-Network, such as the server responding later to a client subscription.

/**
 * This code is used to read the tid and payload from a wire message,
 * and send the same tid and message back to the client
 */
public class WireEchoRequestHandler extends WireTcpHandler {

    public WireEchoRequestHandler(@NotNull Function<Bytes, Wire> bytesToWire) {
        super(bytesToWire);
    }


    /**
     * simply reads the csp,tid and payload and sends back the tid and payload
     *
     * @param inWire  the wire from the client
     * @param outWire the wire to be sent back to the server
     * @param sd      details about this session
     */
    @Override
    protected void process(@NotNull WireIn inWire,
                           @NotNull WireOut outWire,
                           @NotNull SessionDetailsProvider sd) {

        inWire.readDocument(m -> {
            outWire.writeDocument(true, meta -> meta.write(() -> "tid")
                    .int64(inWire.read(() -> "tid").int64()));
        }, d -> {
            outWire.writeDocument(false, data -> data.write(() -> "payloadResponse")
                    .text(inWire.read(() -> "payload").text()));
        });
    }
}

The client code that creates the TcpChannelHub,

The TcpChannelHub is used to send your messages to the server and then read the servers response.

The TcpChannelHub ensures that each response is marshalled back onto the appropriate client thread. It does this through the use of a unique transaction ID ( we call this transaction ID the "tid" ), when the server responds to the client, its expected that the server sends back the tid as the very first field in the message. The TcpChannelHub will look at each message and read the tid, and then marshall the message onto your appropriate client thread.

TcpChannelHub tcpChannelHub = TcpChannelHub(null, eg, WireType.TEXT, "",
    SocketAddressSupplier.uri(desc), false);

in this example we are not implementing fail-over support, so the simple SocketAddressSupplier.uri(desc) is used.

Creates the message the client sends to the server

// the tid must be unique, its reflected back by the server, it must be at the start
// of each message sent from the server to the client. Its use by the client to identify which
// thread will handle this message
final long tid = tcpChannelHub.nextUniqueTransaction(System.currentTimeMillis());

// we will use a text wire backed by a elasticByteBuffer
final Wire wire = new TextWire(Bytes.elasticByteBuffer());

wire.writeDocument(true, w -> w.write(() -> "tid").int64(tid));
wire.writeDocument(false, w -> w.write(() -> "payload").text(expectedMessage));

When you have multiple client threads its important to lock before writing the data to the socket.

tcpChannelHub.lock(() -> tcpChannelHub.writeSocket(wire));

In order that the correct reply can be send to your thread you have to specify the tid

Wire reply = tcpChannelHub.proxyReply(TimeUnit.SECONDS.toMillis(1), tid);

// read the reply and check the result
reply.readDocument(null, data -> {
    final String text = data.read(() -> "payloadResponse").text();
    Assert.assertEquals(expectedMessage, text);
});

eg.stop();
TcpChannelHub.closeAllHubs();
TCPRegistry.reset();
tcpChannelHub.close();

By default the Chronicle-Network server uses a single thread, to process all messages. However, if you wish to dedicate each client connection to its own thread. Then you can change the server threading strategy, to :

-DServerThreadingStrategy= MULTI_THREADED_BUSY_WAITING

see the following enum for more details net.openhft.chronicle.network.ServerThreadingStrategy

This library will require Java 8

The target environment is to support TCP over 10 Gig-E ethernet. In prototype testing, this library has half the latency and support 30% more bandwidth.

A key test is that it shouldn't GC more than once (to allow for warm up) with -mx64m

This comes at the cost of scalability for large number os connections. In this situation, this library should perform at least as well as netty.

Netty has a much wider range of functionality, however it creates some garbage in it's operation (less than using plain NIO Selectors) and isn't designed to support busy waiting which gives up a small but significant delay.