基本服务器占用大量CPU资源-如何缓解？

Question

I'm writing a pretty basic server in C++.

I've ended up with an architecture similar to this:

One thread (the server) is constantly looping over accept() waiting for connections (docs) .

Every connection it finds, it spawns a thread that loops over recv() waiting for something to read, and send() that sends out anything in its queue (recv docs) , (send docs) .

However, I want to be able to (from an external thread) tell either the server or individual connections to essentially kill themselves. The problem is that accept() and recv() both block until they find any activity, so there's no way for them to check if their should_kill_self member is set or anything. So to fix that, I called fcntl(sock_fd, F_SETFL, O_NONBLOCK) on the file descriptors so that they won't.

This way, I have the server constantly looping over accept() and checking should_kill_self , and I have each connection looping over recv() , check if anything in queue to send() , and checking should_kill_self .

So this way, I can signal them to kill themselves at any time, but a side effect I've noticed is that they are taking up huge amounts of processing. I am wondering if with this setup I'm doing something stupid resulting in redundant computation. I don't need this to be a super-low-latency server AT ALL, so is the "correct" way to alleviate this by like, sleeping for 10 seconds between every loop?

I'm new to writing networking code, so I'm pretty sure I'm just doing something stupid. Any advice would be greatly appreciated!

Answer 1

You probably need to explore non-blocking IO in Linux or IO completion ports on windows. See here and here .

Programming these needs extra effort and is abit harder than programming the simple socket server. However, there are libraries that will wrap non-blocking IO (and IOCP) and provides a simple interface for you to use. One example being ASIO in boost. There are many other libraries even for Java and other languages.