线程，事件循环以及大量的连接和并发

Question

all. I'm trying to figure out the best method of handling large amounts of concurrent connections (upwards of more than 800k) in the best manner possible. I've decided to go with libevent for handling reads/writes on the sockets, and will be using one port on the back end and handing the fd around on non-blocking sockets. Where my issue(s) come in to play:

1) libevent with regards to event bases - If I were to, for instances, run one thread per a core each with its own libevent base listening on a single fd for incoming connections, how would libevent handle multiple event bases being triggered on that single fd? The idea would be to then take that incoming connection, accept it and start a new event base on the new fd specific to that one connection. Or, is the appropriate way to do it as I have in the past - run the primary event base in the main program thread and push the incoming connection off to a worker thread to handle the accept, and then create a new event base for each connection then?

2) Threads per a connection... Yes, or no? In past implementations, I've done a model of 1:1 threads per an accepted connection. The end result is obviously with 500 clients connected, 500 threads plus whatever threads I've used as a work queue. However, I'm concerned that this may pose an issue once I reach in to the hundreds of thousands marker connection-wise... Can anyone confirm? I also feel like having that many threads when using something with asynchronous IO such as libevent is not necessary and simply adds more overhead than what's needed... But I could be wrong.

This is the first time I'm having to write something that will support this high of a user load, and I'd prefer to be able to write it from scratch from a solid conceptual design as a means of fully understanding everything at play. Obviously I could go dig through something like UnrealIRCD or nginx's source code and come up with a solution, but I'd really prefer to do it from an understanding of what it is I'm writing and why I'm writing it that way. So some feedback would be greatly appreciated.

Answer 1

I believe you may be confusing two different programming paradigms. If you are using libevent, each thread will handle hundreds or thousands of connections. For more information as to where to begin developing a high volume/concurrent server search for the C10K problem. Here is a page to start:

http://www.kegel.com/c10k.html

BTW, if you want to scale to 800K concurrent connections you should not use a "thread per connection" model.