Netty High performance reason, Netty tuning

From a large perspective, Netty performance is mainly reflected in: 1.io threading Model using reactor mode, synchronous non-blocking. This determines that more can be done with the least amount of resources. 2. Memory 0 copies use direct caching 3. Memory pool design application memory can be reused, mainly direct memory. The internal implementation is to manage memory allocations with a binary lookup tree. 4. The serialization processing socket reads and writes, avoids the lock, namely one specifies the socket the message is the serialization processing. This performance is better than multiple threads processing a single socket counterpart message at the same time, because multithreading can have locks. 5. Provide support for high performance serialization protocols such as Protobuf
Application Netty Tuning: 1. If you only want the handler to finish processing, to the next handler processing, then call: Channelhandlercontext Fireread, and so on. Direct through Channel.read or write will trigger the complete handler processing chain.
The message is automatically released when the 2.SimpleChannelInboundHandler is finished and all handler are processed. Channelinboundhandleradapter does not automatically release messages.
The 3.ServerSocketChannel loop defaults to a maximum of 16 client connections, mainly because the server has a large number of clients that guarantee higher throughput. Corresponding parameter: max_messages_per_read
4. Write_buffer_low_water_mark can be used to do flow control needs to apply call iswritable method, if return false temporarily write failure.
5. Direct memory usage, 0 copies. Direct memory block cache pool design.
6. Avoid multithreading in the memory block cache to get blocks of memory blocking competition, small block of memory size from the thread unique cache to get.
7.reactor mode core is to handle the most things with minimal resources, a reactor thread is responsible for viewing registered care events
8. Soft interrupt Linux 2.6.35 The above version can be set RPS, the main purpose is the processing of packets evenly scattered on each CPU. Improve network processing capabilities. 9. Set TCP parameters is mainly to receive buffer size: SO_RCVBUF, and send buffer size: so_sndbuf. Turn off the So_tcpnodelay for applications that require little delay.