Monday, June 18, 2012

Netty Tutorial Part 1.5: On Channel Handlers and Channel Options

Intro:

After some feedback on Part 1, and being prompted by some stackoverflow questions, I want to expand on and clarify some topics, so this is Part 1.5.
  • Channel Handler Sharability & State
  • Channel Options

Channel Handlers

As discussed previously, most types of ChannelHandlers have the job of either encoding objects into byte arrays (or specifically, ChannelBuffers) or decoding ChannelBuffers into objects. In most cases, handlers will be instantiated and placed into a ChannelPipeline by a defined ChannelPipelineFactory in code that looks something like this (borrowed from the ClientBootstrap javadoc):

The seeming complexity of the pipeline factory may not be obvious, but it's actually an elegant structure that  flexibly handles some variability in the capabilities of different ChannelHandler implementations. First of all, many ChannelHandlers require some configuration, a detail not represented in the above example. Consider the DelimiterBasedFrameDecoder which decodes chunks of bytes in accordance with the constructor specified delimiter. Instances of this decoder require configuration, so the interface only ChannelPipelineFactory allows us to provide the template that defines how it should be created every time a new pipeline is created. An as a recap, every time a new channel is created, a pipeline is created for it, although in some cases, I suppose, you might end up with an empty pipeline with no channel handlers, which means you are using a snazzy framework for passing around ChannelBuffers. That, combined with the ordering and logical names of handlers in the pipeline, is why one needs some care and attention when creating a pipeline. Here's a better example of the use of some parameterized channel handlers in use (again, pilfered liberally from the StringEncoder javadoc):

Shared and Exclusive Channel Handlers

However, a crucial factor in the use of channel handlers is this: Some channel handlers are such good multitaskers, that you only ever need one instance and that single instance can be reused in every single pipeline created. That is to say, these types of channel handlers maintain no state so they are exhibit channel safety, a variation on the notion of thread safety. With no state to maintain, there's no reason to create more than one. On the other hand, other channel handlers keep state in instance variables and therefore assume that they are exclusive to one channel and one pipeline. Both the above examples create pipelines with exclusive channel handler instances, inasmuch as they create a new handler instance each time a new pipeline is created. The second  example, creating a pipeline factory called MyStringServerFactory (which uses actual real-life channel handlers) contains an example of two channel handlers which could actually be shared, although they have been created in exclusive mode. They are the StringEncoder and StringDecoder (and I will explain the reasoning behind it shortly).
Creating and registering a channel handler in shared mode is simply a matter of instantiating one instance and re-using it in all created pipelines. The following code reimplements MyStringServerFactory using a shared StringDecoder and StringEncoder:

So how is one to know if a channel handler class is sharable or exclusive ? Netty defines an informational annotation called @Sharable and sharable channel handler classes are annotated as such. To quote from the @Sharable's javadoc: 
Indicates that the same instance of the annotated ChannelHandler can be added to one or more ChannelPipelines multiple times without a race condition.
If this annotation is not specified, you have to create a new handler instance every time you add it to a pipeline because it has unshared state such as member variables.
This annotation is provided for documentation purpose, just like the JCIP annotations.
It's safe to say that the author summed it up with more brevity than I did, but also informs the reader that the annotation is informational only, so there is no built-in use of the annotation within the API implementation itself, although the @Retention of the annotation is RUNTIME so it can be detected by your code, so using a bit of Reflection[s], I whipped up this table that categorizes all the channel handlers included in the Netty API into Sharable, Exclusive, Upstream, Downstream, and Both. Excluded from this table are abstract and inner classes as well as SimpleChannelHandlerSimpleChannelUpstreamHandler and  SimpleChannelDownstreamHandler. Those last three are concrete channel handlers and are not annotated as @Sharable but they are, what I refer to as, logically abstract, or empty implementations. By themselves, they do nothing and are intended for extending when implementing your own handler classes.
All packages are in org.jboss.netty.
Direction Sharable Exclusive
Upstream
  • handler.codec.string.StringDecoder
  • handler.codec.protobuf.ProtobufDecoder
  • handler.timeout.IdleStateHandler
  • handler.timeout.ReadTimeoutHandler
  • handler.codec.base64.Base64Decoder
  • handler.codec.http.HttpResponseDecoder
  • handler.codec.rtsp.RtspRequestDecoder
  • handler.codec.protobuf.ProtobufVarint32FrameDecoder
  • handler.codec.frame.LengthFieldBasedFrameDecoder
  • handler.codec.frame.DelimiterBasedFrameDecoder
  • handler.codec.rtsp.RtspResponseDecoder
  • handler.codec.spdy.SpdyHttpDecoder
  • handler.codec.serialization.ObjectDecoder
  • handler.codec.http.websocket.WebSocketFrameDecoder
  • handler.timeout.IdleStateAwareChannelUpstreamHandler
  • handler.codec.http.websocketx.WebSocket08FrameDecoder
  • handler.codec.http.HttpChunkAggregator
  • handler.codec.serialization.CompatibleObjectDecoder
  • handler.codec.compression.ZlibDecoder
  • handler.codec.http.websocketx.WebSocket00FrameDecoder
  • handler.codec.spdy.SpdyFrameDecoder
  • handler.codec.http.HttpRequestDecoder
  • handler.codec.frame.FixedLengthFrameDecoder
  • handler.queue.BlockingReadHandler
  • handler.codec.http.HttpContentDecompressor
  • handler.codec.http.websocketx.WebSocket13FrameDecoder
Downstream
  • handler.codec.base64.Base64Encoder
  • handler.timeout.WriteTimeoutHandler
  • handler.codec.rtsp.RtspResponseEncoder
  • handler.codec.protobuf.ProtobufEncoder
  • handler.codec.protobuf.ProtobufVarint32LengthFieldPrepender
  • handler.codec.http.websocket.WebSocketFrameEncoder
  • handler.codec.http.websocketx.WebSocket00FrameEncoder
  • handler.codec.rtsp.RtspRequestEncoder
  • handler.codec.serialization.ObjectEncoder
  • handler.codec.frame.LengthFieldPrepender
  • handler.codec.string.StringEncoder
  • handler.codec.serialization.CompatibleObjectEncoder
  • handler.codec.http.HttpResponseEncoder
  • handler.codec.compression.ZlibEncoder
  • handler.codec.spdy.SpdyFrameEncoder
  • handler.codec.spdy.SpdyHttpEncoder
  • handler.codec.http.websocketx.WebSocket13FrameEncoder
  • handler.codec.http.websocketx.WebSocket08FrameEncoder
  • handler.codec.http.HttpRequestEncoder
Both
  • handler.logging.LoggingHandler
  • handler.execution.ExecutionHandler
  • handler.timeout.IdleStateAwareChannelHandler
  • handler.queue.BufferedWriteHandler
  • handler.codec.spdy.SpdyFrameCodec
  • handler.codec.spdy.SpdySessionHandler
  • handler.codec.spdy.SpdyHttpCodec
  • handler.ssl.SslHandler
  • handler.stream.ChunkedWriteHandler
  • handler.codec.http.HttpClientCodec
  • handler.codec.http.HttpContentCompressor
  • handler.codec.http.HttpServerCodec

Decoders

You might notice a few instances of asymetry when it comes to sharability. For example, the ObjectEncoder is sharable, but the ObjectDecoder is not. This is revealing about the nature of handlers. An ObjectEncoder's job is to convert an object into a byte array (ok, a ChannelBuffer), and it is passed the complete payload (the object to be encoded) so there is no need to retain any state. In contrast, the ObjectDecoder has to work with a bunch of bytes which may, or may not be complete for the purposes of reconstituting back into an object. Think about a serialized object being sent from a remote client. It has been serialized into a byte stream of possibly several thousand bytes. When the decoder is first called, the bytestream may be incomplete, with some portion of the byte stream still being transported.
Don't even think about having the handler wait for the rest of the byte stream to arrive before proceeding. That sort of thing is just not done around these parts, because Netty is an asynchronous API and we're not going to allocate a pool of worker threads (remmeber, they do all real work) to sit around waiting for I/O to complete. Accordingly, when there are bytes available for processing, a worker thread will be allocated to process it through execution of the ObjectDecoder's decode method. This is the signature of that method:

Object decode(ChannelHandlerContext ctx, Channel channel, ChannelBuffer buffer)
 
When this method is called, if the buffer contains the complete set of bytes required to decode into an object, then the decoded object will be returned and passed on to the next upstream handler. If not, the handler returns a null, in effect signaling a No Op. Either way, the worker thread doing the work is off onto doing something else once the invocation is complete. 
 
That's a standard pattern for decoders. Examining ObjectDecoder specifically, we can see that it extends LengthFieldBasedFrameDecoder, and a brief glance at the source of that class shows the following non-final instance fields:
 

private boolean discardingTooLongFrame;
private long tooLongFrameLength;
private long bytesToDiscard;
 
These fields are the state being protected by making this handler exclusive. The actual byte stream being  decoded is not stored in state, but rather accumulates in the same instance of the ChannelBuffer passed to the decode each time an invocation is made to decode the specific byte stream. This is an important conceptual point to understand: A decoder may be invoked more than once in order to decode a byte stream. Completion results in a returned object;  in-completion results a returned null (and an assumption of more invocations to follow). 

Non-Blocking = Scalability ?

An important takeaway from this is the notion that worker threads only kick in when there is work to do, rather than blocking waiting for I/O to deliver bytes to process. Consider a simple and contrived scenario where a server listens for clients submitting serialized java objects. Implementing a blocking I/O reader, the thread spawned to process a given client's object submission might create a java.io.ObjectInputStream to read from the SocketInputStream. Bytes are read in and objects are spit out the other side. If the byte stream is constantly full, then we would have efficient use of worker threads, although the model still requires one thread to be dedicated to each connected client. Fair enough, since there is a constant supply of data to process, all things being equal.




However, what happens if the supply of bytes to process is intermitent ? Perhaps the clients only supply data intermitently, or perhaps the network has limited bandwidth and/or is congested, delaying the object transmissions. Now the server has allocated threads to read (and block) from the input stream, but there are long periods of time where there is nothing to do but wait for work to arrive.




This is one of the reasons that non-blocking I/O tends to scale better than traditional blocking I/O. Note, it does not necessarilly make it faster, it allows the server to accept and process an escalating number of clients for a more controlled and reduced level of resource allocation. For smaller numbers of clients,  it is quite likely that blocking I/O will be faster, since dedicated threads require less supervision. Of course, since Netty supports both blocking (OIO) and non-blocking (NIO) channels, a Netty developer can create all sorts of services that can be configured to use either. Which one is appropriate for any given application is so dependent on many application specific factors that it is rarely possible to state any hard rules about this. The Netty javadoc for the org.jboss.netty.channel.socket.oio package attempts to set a rule of thumb regarding the use of OIO:
Old blocking I/O based socket channel API implementation - recommended for a small number of connections (< 1000).
That seems high to me, but hardware, network, traffic types and other factors will certainly influence the magic number.

 Framing  (buffers, not pictures)

 ObjectDecoder extends LengthFieldBasedFrameDecoder because the latter performs a critical function necessary in many decoders. Without a frame decoder, the object decoder would be faced with an entire unsegmented byte stream, not knowing where one serialized object ended and the next one started. 









 
Of course, the CompatibleObjectDecoder can do this because it expects a standard java serialization stream, but there are several downsides to this, not the least of which is that they payloads will be significantly larger. In order to help the ObjectDecoder figure out which bytes belong to which serialized object, the frame decoder part of the class knows how to segment the byte stream into discrete chunks that the ObjectDecoder can assume are the contents of one object.











The frame decoder knows how to segment the byte stream because it works in concert with the sender's ObjectEncoder, which has conveniently placed length indicators into the downstream byte stream, specifically so the upstream decoder can use them to parse. When implementing your own custom decoders, if you wanted to implement this technique, your client can use the LengthFieldPrepender encoder.

There are two points to note from this:

  • Most non-trivial decoders are either implementing a known protocol specification (like HTTP for example) or are expecting the payload to have been "massaged" by the encoding counterpart (like the ObjectEncoder).
  • The opportunity for optimization with the Netty encoder/decoder pairs is a good (though not the only) reason to use Netty even if you're using the traditional blocking IO ( OIO ) model.

In the case of the ObjectDecoder, it has it's own frame decoding built in, but in many decoder classes, there is no frame decoding built in and framing support needs to be provided. Fortunately, the modular architecture of the ChannelPipeline allows us to insert a frame decoder before instances of decoders that need them. A good example of this is the StringDecoder. The javadoc emphasizes that this decoder should be used in conjunction with a frame decoder:

Please note that this decoder must be used with a proper FrameDecoder such as DelimiterBasedFrameDecoder if you are using a stream-based transport such as TCP/IP.

There are a few ways that transmitted strings might be represented and the corresponding frame encoder and decoder that can be used:
  1. Fixed Length:   No specific encoder, just use fixed length strings / FixedLengthFrameDecoder
  2. Length Field Prepend: LengthFieldPrepender / LengthFieldBasedFrameDecoder
  3. String delimiter: No specific encoder, but strings should be delimited by the specified delimiters / DelimiterBasedFrameDecoder
The DelimiterBasedFrameDecoder can be configured with more than one delimeter and the presence of either in the stream will trigger a frame segmentation. Typical string delimiters are end-of-line characters or null characters. Delimiters are specified not as strings or even characters, but as ChannelBuffers so they are expressed using the same type that the decoder's input and output will be in.

The class Delimiters contains a few pre-defined and commonly used delimiters, and it's simple enough to define your own:

ChannelBuffer delimiter = ChannelBuffers.wrappedBuffer("THEBIG-D".getBytes())


To frame-up this discussion on frames, channel handlers are often put to work in conjunction with each other. This can be within one pipleline, such as a frame decoder before a string decoder. In addition, a decoder's decoding algorithm may rely on a counterpart encoder's encoding.

ChannelHandler State


Although the most common (and recommended) way of handling state in a ChannelHandler is to use instance fields and then make sure the handler is used exclusively ( a new instance per pipeline inserted into ), there are a couple of ways of keeping state safely in a shared ChannelHandler.

The ChannelHandlerContext class provides the notion of an attachment which is a store that is specific to a channel and handler instance. It is accessed using the getter and setter methods:

void channelHandlerContext.setAttachment(Object obj)
Object channelHandlerContext.getAttachment()

Typically, if you only need to keep state within a handler across calls to the handler, the attachment model is suitable. However, if you need to access state specific to a channel outside [and/or inside] the handler, Netty provides an analog to ThreadLocal called a ChannelLocal. Obviously, since this state is channel specific, but may be accessed by multiple different threads, a ThreadLocal does not serve this purpose. While a ThreadLocal is essentially a bit of data keyed in part by the owning thread, a ChannelLocal is similar in that it is a bit of data keyed by the owning channel. Accordingly, you can define a static field and using the channel as the key, access the value from pretty much anywhere in your code. The primary methods are:


void channelLocal.set(Channel channel, T value)
T channelLocal.get(Channel channel)


This concludes the discussion specific to handlers. It seems to me that a missing piece here is a walk-through of building a read world encoder/decoder pair that really does something useful, and I will endeavor to do this in a future entry.

Channel Options


Channel options are configuration name/value pairs that are used as a loosely typed way of configuring various bits-and-pieces in Netty. They fall into four categories:


  • General Netty Configuration
  • Client Socket Configuration
  • Server Socket Configuration 
  • Server Child Socket Configuration 

In essence, channel options are placed in a Map and then passed to the Bootstrap; the options provided are applied where applicable, so this is an abstracted way of configuring the stack.

General Netty Channel Options

There are a number of option keys that can be used in place of specific API calls to configure how channels are created. The following is a partial summary of the keys and what the types are just to give you an idea what they are. 

  • bufferFactory: An instance of a ChannelBufferFactory that will be called to create new ChannelBuffers. This might be useful in order to customize the byte order of created channels, or to create direct (off heap) buffers.
  • connectTimeoutMillis: The number of milliseconds before a connection request times out. Ignored if the channel factory creates connectionless channels.
  • pipelineFactory: An instance of a ChannelPipelineFactory
  • receiveBufferSizePredictor: UDP packets are received into pre-created byte arrays, which is an unresolvable dilema in that you don't know how much space to allocate until you have received the datagram, but to receive the datagram, you have to allocate a buffer to receive it. BufferSizePredictors provide a way to customize the size of those arrays based on some intelligence you might have about the expected size.
  • receiveBufferSizePredictorFactory: A factory for creating buffer size predictors.

Socket Channel Options

Most of the supported channel options are for configuring sockets. By and large, Netty does not really expose the Java Socket API that much, which might seem odd for a networking API. However, all implementations of SocketChannel use sockets, and sockets have various low level configuration items which can be set through channel options. In general, these are the types of sockets we are working with:

  • Client Socket: A socket on the client side of a TCP connection to a server.
  • Datagram Socket: A socket on the client and/or server side of a unicast UDP conversation.
  • Multicast Socket: A socket on the client and/or server side of a multicast UDP conversation. 
  • Server Socket: A specialized socket on the server side of a TCP server that brokers connection requests from client sockets and spins up a new Socket dedicated to that connection in an action referred to as an accept.  Aside from the initial connection, a client socket does not communicate with the server socket.
  • Child Socket: The socket created on the server, by the server socket, to communicate with a client socket.

All of these socket types support options outlined in java.net.SocketOptions and many of them can have a significant impact on performance. The options are listed below with the SocketOption constant name and the Netty channel option equivalent.



  • IP_MULTICAST_IF / networkInterface: The name of the interface on which outgoing multicast packets should be sent. When a host has multiple network interfaces, this tends to be quite important.
  • IP_MULTICAST_IF2 / networkInterface: The same as IP_MULTICAST_IF but defined again for good measure.
  • IP_MULTICAST_LOOP / loopbackModeDisabled: Defines if multicast packets should be received by the sender of the same.
  • IP_TOS / trafficClass: Sets the type of service or traffic class in the headers of packets.
  • SO_BINDADDR / : A read only option representing the bind interface of a local socket.
  • SO_BROADCAST / broadcast: Enables or disables a DataGramSocket's ability to send broadcast messages,
  • SO_KEEPALIVE / keepAlive: A flag indicating that probes should be periodically sent across the network to the oposing socket to keep the connection alive.
  • SO_LINGER / soLinger: Allows the close of a socket to be delayed for a period of time to allow completion of I/O.
  • SO_OOBINLINE / : Allows Out of Band TCP Urgent data transmissions to be received through the normal data channel rather than discarded.
  • SO_RCVBUF / receiveBufferSize: The size in bytes of the socket's data receive buffer.
  • SO_REUSEADDR / reuseAddress: When set to true, allows multiple DatagramSockets to be bound to the same socket address if the option is enabled prior to binding the socket.
  • SO_SNDBUF / sendBufferSize: The size in bytes of the socket's data send buffer.
  • SO_TIMEOUT / connectTimeoutMillis: Sets a timeout in ms. on blocking socket operations. 
  • TCP_NODELAY / tcpNoDelay: Disables Nagle's algorithm on the socket which delays the transmission of data until a certain volume of pending data has accumulated.

SO_Timeout vs. ChannelFuture Await

The connectTimeoutMillis (or Socket SO_TIMEOUT) is sometimes confused with the timeout that can be set on a ChannelFuture using await or one of its equivalents. The SO_TIMEOUT is strictly the time limit (in ms.) that an I/O operation on the socket must complete in. If a connect I/O operation is not complete within that time, the socket instance will throw a java.net.ConnectException: connection timed out exception. On the other hand, the ChannelFuture's await timeout is simply the amount of time the calling thread will wait for the operation to complete. The timeout itself does not impact the operation and the operation is considered complete when it concludes, whether it was successful or not.

Having said that, if you wanted any asynchronous operation executed on a channel to be cancelled if it did not complete within a specific period of time, a ChannelFuture await timeout is useful because when the timeout elapses and the operation is found incomplete, it can be cancelled. This might be applicable, for example, in a mobile device scenario where you might want to save battery consumption by ensuring that network operations did not run an overly long time. Just because your application threads are not running, does not mean the device would not be consuming power attempting to complete a long running network operation that the application has long since forgotten about.

The following example will attempt to illustrate the difference and how they can be usefully implemented. This code sample is based on simple testing framework class called SimpleNIOClient. The code for the example is TimeoutTest. The test performed is very simple and just attempts to connect to an internet address using different values for the ChannelFuture await timeout and the client socket's SO_TIMEOUT.

The salient part of this test is the connect code and the timers that are placed around it.

 public static void timeoutTest(long ioTimeout, long futureTimeout) {
  // Create a map with the channel options
  Map<String, Object> options = new HashMap<String, Object>();
  options.put("connectTimeoutMillis", ioTimeout);
  // Create the client
  // Not providing any handlers since we're not using any for this test
  SimpleNIOClient client = new SimpleNIOClient(options);
  // Issue a connect operation
  ChannelFuture cf = client.connect("heliosapm.com", 80);  
  // Add a completion listener
  cf.addListener(new ChannelFutureListener() {
   public void operationComplete(ChannelFuture future) throws Exception {
    if(future.isSuccess()) {
     clog("F: Connected:" + future.isDone());
    } else {
     if(future.isCancelled()) {
      clog("Request Cancelled");
     } else {
      clog("Connect Exception:Success: " + future.isSuccess() + "  Done: " + future.isDone()  + "  Cause: "+ future.getCause());
     }
    }
   }
  });
  // Wait at least futureTimeout for the operation to complete
  cf.awaitUninterruptibly(futureTimeout);
  // If the operation is not complete, cancel it.
  if(!cf.isDone()) {
   clog("Channel Future Still Waiting. Cancelled:" + cf.cancel());
  }
 }

I run the test using arbitrarily low timeouts since I am trying to trigger connect timeouts. The invocation code looks like this:
  // Timeout on I/O  
  timeoutTest(10, 500);
  // Wait a bit so output is clear
  try { Thread.sleep(2000); } catch (Exception e) {}
  System.out.println("===============================");
  // Timeout on future
  timeoutTest(500, 10);
  try { Thread.currentThread().join(5000); } catch (Exception e) {}

In the first case, on line 2, I want the SO_TIMEOUT timeout to occur, so I invoke with a low I/O timeout and a higher ChannelFuture timeout. On line 7, I reverse the values for a higher timeout on the socket, but a low timeout on the ChannelFuture. The output is as follows:


[Client][Thread[ClientBossPoolThread#1,5,ClientBossPoolThreadGroup]]:Socket Timeout Exception:Success: false  Done: true  Cause: java.net.ConnectException: connection timed out
===============================
[Client][Thread[main,5,main]]:Request Cancelled
[Client][Thread[main,5,main]]:Channel Future Still Waiting. Cancelled:true


In the first case, the socket timeout is so low that the operation completes quite quickly (although it does not succeed) so the future listener reports that the operation:

  1. Was not successful
  2. Is Complete
  3. Resulted in a ConnectException on account of a timed out connection.
In the second call, the higher socket timeout means the future timeout expires first, and when it does, the operation is still running, so the thread cancels it.

Socket Buffer Sizes


Another critical performance affecting channel option are the socket buffer sizes, which are the sizes of the send and receive native platform buffers allocated by the networking kernel to buffer I/O. It is possible to simply set these values arbitrarily high and leave it at that, but in situations where there are a large number of sockets in use, this may waste native memory resources.

The next example code I want to demonstrate uses another test harness called SimpleNIOServer. The test class is called AdjustableSocketBufferSizeServer. It requires a little explanation around some of the customizations, so please bear with....

ChannelHandlerProviderFactory

This is a channel handler factory that understands whether a channel handler class is sharable or exclusive, so when an instance of a ChannelHandlerProvider is given to the ChannelPipelineFactoryImpl and it creates a new pipeline, it will either create a brand new handler instance (if exclusive) or provide the "singleton" instance already created. This is not critical to the example, but to avoid confusion, I am mentioning their use.

CustomServerBootstrap

This is a customized server bootstrap that allows me to change the channel options on the fly. In the native version, once a bootstrap has been created, the channel options are committed and I want to to be able to change the socket buffer sizes of created child channels on the fly. (Once a change is made, it affects only new connections from that point on.... it does not change existing socket buffer sizes). I also created a management interface called BootstrapJMXManager that allows me to make these changes through JMX.

Alright, the test class, AdjustableSocketBufferSizeServer, creates a new server using SimpleNIOServer. It also creates a JMXConnectorServer and RMI Registry so the MBeanServer can easily be connected to so the child socket receive buffer sizes can be adjusted programatically, so that in turn the test case can be run automatically end to end.

The server's pipeline factory creates pipelines with the following handlers:

  • Upstream: 
    • InstrumentedDelimiterBasedFrameDecoder with a "|" delimiter (exclusive)
    • StringDecoder (shared)
    • StringReporter (shared)
  • Downstream:
    • CompatibleObjectEncoder (shared) 

The test is fairly straightforward:
  1. The client makes a JMX call to the server to specify the size of the server allocated child channel's socket receive buffer size.
  2. The client makes a new connection and sends a [longish] string to the server.
  3. The FrameDecoder splits the string content and sends the fragments upstream to the StringDecoder.
  4. The StringDecoder decodes the byte stream into a String and sends the strings upstream to the StringReporter.
  5. The StringReporter measures the length of the String, creates an int array with the length of the passed string and writes the array downstream back to the client.
  6. On its way downstream, the int[] is encoded using standard Java serialization in the CompatibleObjectEncoder.
  7. The client records the elapsed time to get a response from the server.
  8. The client increases the requested server allocated child channel's socket receive buffer size.
  9. Rinse, Repeat, (Go to #1.) for each subsequent defined buffer size
The script looping has an outer loop and an inner loop:

  • Outer Loop: Iterates each socket receive buffer size defined in an array, sets this size on the server's bootstrap and executes the inner loops. Size: 5
  • Inner Loop: Submits the configured string payload and captures the elapsed time to response. Size: 100


I am going to run the client side of the test from a Groovy script using the raw TCP Socket API, so the server pipeline uses a CompatibleObjectEncoder since the client cannot decode the optimized ObjectDecoder's payload. (It could, but I am sticking to a generic client).


InstrumentedDelimiterBasedFrameDecoder


Remember that for each logical request, server side decoders may need to be invoked several times in order to complete the task of reading the full bytestream and decoding it. When it is invoked, if the bytestream is not complete, it returns a null. Eventually, the bytestream will be complete and the decoder can complete the task and return the (non-null) decoded payload. This being the case, my theory was that the smaller the receiving socket's receive buffer size, the more times the decoder would need to be called. Or in contrast, the larger the receiving buffer size, the more likely that the full bytestream would be transmitted in one shot, reducing the number of required frame decoder calls to one.


To test this theory, I made a simple extension of the DelimiterBasedFrameDecoder that counts the number of times it is invoked by incrementing a counter in an instance variable. When the decode completes, the highwater counter value is written to a ChannelLocal and the counter is reset back to zero. Upstream, the StringReporter, which determines the length of the passed string and returns it to the client, also returns the count of frame decoder invocations by reading it from the ChannelLocal and adding it to the int[] payload returned to the client. In addition, it occurred to me that the socket buffer size might not actually be set to what the client requested. (Will it really set the buffer size to 2 bytes ??!!??) So I added a read of that value and appended it to the result sent back to the client. It is useful to note that when reading or writing channel options from a live (socket based) channel, they are read from and applied to  the actual socket instance. With that, the client can report:

  • The length of the submitted string reported by the server. (Not useful except as a validating assertion)
  • The number of times the frame decoder was called for the most recently submitted string.
  • The elapsed time of the request execution.
  • The actual server socket receive buffer size.



The server pipeline uses a CompatibleObjectEncoder since we're sending back an int[] and since the client is a simple socket client, not a Netty based client. It is a bit of overkill though. Whenever one is sending primitives, or arrays of primitives, it is possible to bypass modular encoding and simply write directly to a ChannelBuffer. Here's how we could remove the CompatibleObjectEncoder from the pipeline and rewrite the StringReporter's write of the int[].

   
   // =====================================================================
   //  Uncomment next section to ditch the CompatibleObjectEncoder from the pipeline
   // =====================================================================
   ChannelBuffer intBuffer = ChannelBuffers.buffer(12);
   for(int i = 0; i < 3; i++) {
    intBuffer.writeInt(ret[i]);
   }
   ctx.sendDownstream(new DownstreamMessageEvent(e.getChannel(), Channels.future(e.getChannel()), intBuffer, e.getChannel().getRemoteAddress()));
   // =====================================================================
   
   // =====================================================================
   // Comment the next line to ditch the CompatibleObjectEncoder from the pipeline
   // =====================================================================
//   ctx.sendDownstream(new DownstreamMessageEvent(e.getChannel(), Channels.future(e.getChannel()), ret, e.getChannel().getRemoteAddress()));



Having done that, the groovy script client (coming up next) which normally uses an ObjectInputStream to read objects from the SocketInputStream would be replaced with a DataInputStream and the code would look like this:

                    /*
                     * The prior code
                    if(ois==null) {
                        ois = new ObjectInputStream(socket.getInputStream());
                    }
                    result = ois.readObject();
                     */
                    if(ois==null) {
                        ois = new DataInputStream(socket.getInputStream());
                    }
                    result = new int[3];
                    for(index in 0..2) {
                        result[index] = ois.readInt();
                    }


Groovy Client Script and Network Kernel Tuning 


This is the groovy test client I used to test the server. I admit that this script is a bit messy, but it does the job. I won't go over it in too much detail, excepting drawing your attention to the following:


  • The client sends the same string each time, since the varying value being adjusted to measure the effect is the server's socket receive buffer size. 
  • The string content is fairly arbitrary and generated by appending all the JVM's system property key/value pairs a couple of times. Once generated, the same string is used in each call. 
  • The socket receive buffer sizes tested are defined in this int array: [1144, 2288, 4576, 9152, 18304, 36608, 73216, 146432, 292864, 585728] 


 Since the script assertively validates that the socket receive buffer size is actually what the client set it to, I found that when I requested a buffer size of 146,432, I got an assertion failure with the actual reported buffer size being only 131,071. As it happens, my Linux network kernel configuration had limited the maximum size of of socket receive buffers to 131,071. In order to raise this limit, running in a sudo'ed shell, I did this:


  • sysctl -a | grep mem > ~/sysctl-mem.txt (Saved the current settings for rollback purposes) 
  • sysctl -w net.core.rmem_max=8388608 (Set the maximum buffer size to 8.3 MB) 
  • sysctl net.core.rmem_max (Verified the value was set. Output was "net.core.rmem_max = 8388608") 
  • sysctl -w net.ipv4.route.flush=1 (Flushed the settings so they took effect immediately) 


This is a useful reference for tuning the network kernel in Linux. Here is a guide for TCP tuning on different operating systems. I should also note that the string I generated as the sample test payload was 8,088 bytes before being delimited with a single "|".

Here are the charted results of the test, with the socket receive buffer size on the x-axis and the average elapsed time of each request in ms. on the first y-axis. The second y-axis represents the average number of times the Frame Decoder was called for each request.



The results are pretty much what I expected and I think quite intuitive:

  • The average elapsed time of the request executions is inversely proportional to the server child socket's receive buffer size until the buffer size is large enough to contain the whole request payload.
  • The average number of frame decoder invocations per request execution is inversely proportional to the server child socket's receive buffer size until the buffer size is large enough to contain the whole request payload.
One might be tempted to simply set the socket buffer size to the max, but keep in mind that on a busy system, many sockets may be allocated and even though the buffer memory does not come out of your Java heap space, it all adds up in your native memory and could be quite wasteful. Furthermore, it is quite obvious from the chart that for a predictable payload size, there is a sweet spot for buffer sizes, so in concert with Netty's ReceiveBufferSizePredictor, it would be interesting to create an auto-tuning socket buffer sizing algorithm, but that's for another day.

The spreadsheet I used for this test is here, and is simple to use. The groovy test script generates a CSV file with all the tabulated results, so you can just import it into OpenOffice (et. al.) and then paste the numbers in.

That's it for entry 1.5. I hope it has been helpful, if not outright entertaining. I am delighted to get any feedback, and genuinely pleased to answer any questions. As promised, in entry 2.0, I will be discussing implementing Ajax push in Netty (with no further decimal releases in the 1 range).

Part 1 of this series:  Netty Tutorial Part 1: Introduction to Netty


12 comments:

Anonymous said...

Wow~!! I appreciate for your post!! :)

Anonymous said...

awesome!!!!
i'm starting to write a server with netty and i'm so glad that there're your tutorials. it helps a lot to get up an running!

-Venkat said...

Really helpful. Appreciate all your effort in writing such a detailed post. Hope to see more....

Anonymous said...

The article is so good, appreciate all your effort!

Joe.S said...

Very nice,thx very much, waiting for part 2 !

danicricco said...

awesome blog! You managed not only to transmit the concept, but you also have given very useful information

elfogre said...

Both of your articles about Netty are really helpful!

Thank you very much.


*PD: Maybe has a little mistake your RetryingDecoder.png image? On downstream there is a CompatibleObjectDecoder channelHandler instead of a CompatibleObjectEncoder.

David Komer said...

Great articles! Quick question- is there any way to reliably detect when the channel has been closed due to user disconnect? I'm extending ChannelInboundMessageHandlerAdapter in case that helps.

I've also seen a couple StackOverflow points on this but they're only addressing 3.x specifically...

s2o said...

thank you very much for your post!!

egor said...

nickman, thank you for your posts on netty. I think jboss team should reference them as easy to follow tutorial. I was trying to follow their docs on site, but your posts are much easier to understand.

Andrey Cizov said...

Please note that the link to the previous version of the article is incorrect.

nickman said...

Thanks Andrew. I thought I fixed it a couple of times, but something wacky was up with blogger's composer. Finally I went into the HTML editor and it was all dreadful ! Fixed now.