latency in netty due to passing requests from boss thread to worker thread?
I have some questions about Netty (Server Side), TCP/IP applications;
I am wondering if there can be latency because of netty (due to missing configuration etc.) while passing the request from boss thread to worker thread ?
I am using :
new OrderedMemoryAwareThreadPoolExecutor(350, 0, 0, 1, TimeUnit.SECONDS);
Actually, I set max thread count 350 as I am not sure about the optimal number. I log simultaneous working thread count every minute and it seems that average is too low (barely exceeds 10 ). So I will decrease this number as it is not required.
Is there any other parameters,important points that I should be aware of for to get best performance ?
bootstrap.setOption("tcpNoDelay", true); - Is there any disadvantage of setting this parameter? Considering that delivery time is very important.
Thread Pool Executer:
OrderedMemoryAwareThreadPoolExecutor executor = new OrderedMemoryAwareThreadPoolExecutor(48, 0, 0, 1, TimeUnit.SECONDS);
Here is my pipeline factory:
ChannelPipeline pipeline = pipeline();
pipeline.addLast("frameDecoder", new DelimiterBasedFrameDecoder(GProperties.getIntProperty("decoder.maxFrameLength", 8000 * 1024), Delimiters.nulDelimiter()));
pipeline.addLast("stringDecoder", new StringDecoder( CharsetUtil.UTF_8 ));
pipeline.addLast("frameEncoder", new NullTermMessageEncoder());
pipeline.addLast("stringEncoder", new JSONEncoder( CharsetUtil.UTF_8 ));
pipeline.addLast("timeout", new IdleStateHandler(idleTimer, 42 , 0, 0));
pipeline.addLast("executor", new ExecutionHandler(executor));
pipeline.addLast("handler", new GServerHandler());
and the ServerBootstrap:
gServerBootstrap = new ServerBootstrap(new NioServerSocketChannelFactory(Executors.newCachedThreadPool(), Executors.newCachedThreadPool()));
gServerBootstrap.setPipelineFactory(new GServerPipelineFactory());
gServerBootstrap.setOption("backlog", 8129);
gServerBootstrap.setOption("child.tcpNoDelay", true);
gServerBootstrap.bind(new InetSocketAddress(GProperties.getIntProperty("server.port", 7679)));
What can you suggest for this configuration ?
Netty Boss threads are only used to setup connection, worker threads are used to run NioWorker (non blocking read/write) or OioWorker (blocking read/write).
If you have an execution handler, worker thread will submit the message event to OrderedMemoryAwareThreadPoolExecutor.
1) Increasing the Netty I/O worker thread count to more than number of processors * 2 won't help. If you are using staged executors, Having more than one staged execution handler for non I/O tasks, may increase latency.
Note: Its better to set your own ObjectSizeEstimator implementation in OMTPE constructor, because many CPU cycles are spent on calculating used channel memory.
2) There are some other Netty parameters you can try
//setting buffer size can improve I/O
bootstrap.setOption("child.sendBufferSize", 1048576);
bootstrap.setOption("child.receiveBufferSize", 1048576);
// better to have an receive buffer predictor
bootstrap.setOption("receiveBufferSizePredictorFactory", new AdaptiveReceiveBufferSizePredictorFactory(MIN_PACKET_SIZE, INITIAL_PACKET_SIZE, MAX_PACKET_SIZE))
//if the server is sending 1000 messages per sec, optimum write buffer water marks will
//prevent unnecessary throttling, Check NioSocketChannelConfig doc
bootstrap.setOption("writeBufferLowWaterMark", 32 * 1024);
bootstrap.setOption("writeBufferHighWaterMark", 64 * 1024);
3) It should be bootstrap.setOption("child.tcpNoDelay", true) for server bootstrap.
There is an experimental hidden parameter:
Netty NioWorker is using SelectorUtil.select to wait for selector events, the wait time is hard coded in SelectorUtil,
selector.select(500);
setting a small value gave better performance in netty sctp transport implementation. Not sure about TCP.
Q1) Everything adds latency. Netty is pretty efficient so I would be surprised if the latency is too high for 95%+ of use cases
Q2) Test you performance yourself and determine if it a problem, (either latency or throughput) before you start worrying about it.
Q3) This option may help. It shouldn't make it worse. Many modern OSes self tune pretty well and I don't find it makes as much difference as it used to.
Can you clarify what latency you are trying to achieve because it can make a big difference to your design? Is it 10 ms, 1 ms, 100 us, 10 us?
1) I am wondering if there can be latency because of netty (due to missing configuration etc.) while passing the request from boss thread to worker thread ?
I don't think there's much if any latency here. The threads are in a pool, they just need to be given the work.
QUESTION 2) Is there any other parameters,important points that I should be aware of for to get best performance ?
When it comes to 'best' performance, I did a bunch of testing and ended up using a number of threads that was about 16* the number of physical processors on the box. I tried thread numbers up to several thousand, but when they got really slammed, they ended up thrashing in GC.
QUESTION 3) bootstrap.setOption("tcpNoDelay", true);. Is there any disadvantage of setting this parameter ? Considering that delivery time is very important.
Definitely set this.
链接地址: http://www.djcxy.com/p/95210.html上一篇: Java Netty负载测试问题