在研究Java AIO的时候理解到reactor和proactor模式,顺便研究了一下。
这里面会提到blokcing/non-blocking, synchronous/asynchronous的对比,要仔细体会才行。
如果实现一个高性能的服务器端,比如web server,有以下几种方式:
详细的可以参考http://www.cs.wustl.edu/~schmidt/PDF/Reactor1-93.pdf
1、Non-blocking I/O solution
One method for handling I/O on multiple descriptors involves the use of "polling." polling operates by cycling through a set of open descriptors, checking each one for pending I/O activity.
The primary disadvantage with polling is that it consumes excessive CPU cycles by making unnecessary system calls while "busy-waiting". For instance, if input occurs only intermittently on the I/O descriptors, the server process will repeatedly and superfluously poll descriptors that do not have any pending logging records. On the other hand, if I/O is continuously received up all descriptors, this approach may be reasonable. In addition, an advantage with polling is that it is portable accross OS platforms.
2、Multi-Process Solution - master/slave模式
有一个master进程负责监听client请求,然后对于每个请求调用fork方法创建一个单独的进程去处理。
3、Multi-threaded solution
和2类似,只不过将process换成了更轻量级的thread
4、The Event Demultiplexing Solution
主要是用到了操作系统的事件多路分用(event demultiplexing)功能:通过调用select和poll。这一类里根据实现的不同又分为几类:
4.1 select-based
4.2 poll-based
以上两种存在着以下一些缺点:
-> Complicated and Error-Prone Interfaces
-> Low-Level Interfaces
-> Non-Portable Interfaces
-> Non-Extensible Interfaces
为了解决这些问题,我们在更高层面引入了模式,把这些底层的信息屏蔽掉了,使得更易于使用。
4.3 reactive event dispatching model
4.4 proactive event dispatching model
Java AIO的实现者Alan Bateman解释以上两种模式的区别是:
proactive:
-> Initiate non-blocking I/O operation
-> Notification when I/O completes
-> Proactor pattern
reactive:
-> notification when channel ready for I/O (S elector)
-> perform non-blocking I/O operation
-> Reactor pattern
说实话,还不是很能理解其中的奥秘。