Re: [PATCH] Improve scalability of epoll_ctl

[Davide Libenzi]

On Tue, 8 Jan 2008, Eric Dumazet wrote:

> Changli Gao a écrit :
> > Replace the epitem rbtree with a dynamic array to get the constant insertion/deletion/modification time of the file descriptors. Reuse the size argument of epoll_create, however the size must be smaller than the max file descriptor number: ether the resource limitation or the compiling time limitation.
> >
> >   
> Hum, you should read this :
> 
> http://www.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.7-rc3/2.6.7-rc3-mm2/broken-out/epoll-uses-rbtrees.patch
> 
> 
> Your patch is a revert of this change, it probably wont be accepted.
> 
> epoll_ctl() is scalable, since it's O(log2(N)) : With 1 millions files
> descriptors, that means less than 20 nodes to lookup in the tree.

Indeed, and we aren't going back.


> In what situation do you think epoll_ctl() performance is bad ?

I'm sure you can cook up a microbench that only does epoll_ctl(), and with 
a good hash implementation you get better numbers. This is what I did 
before I changed to rbtree, and the numbers did not justify the code (the 
"size" in epoll_create() is just an hint, and that means you need to 
handle resizing - plus other things like locking up memory when the number 
of fds shrinks, and other stuff I don't even remember).



- Davide