2005-02-27 11:05:25

by Neil

[permalink] [raw]
Subject: Yet another I/O modeling

hi,all
select/poll I/O modeling is fast,easy to use,it returns the number of fd
which are acitvity,but you should scan the fd array to find which fd is
ready,it costs a lot of time on a heavy load server.

I proposal another I/O modeling,named eselect.
struct eselect_struct{
unsigned long readbitmap[MAXBITMAPFD];//suppose the MAXBITMAPFD is
MAX OPEN //FD NUMBER PER PROCESS
int ret;//-1 on error,non-negative return value is the number of
//acitvity fds
}
we can use __set_bit(),__clear_bit(),__find_first_bit()....functions to
maintain the bitmap.
So,you shouldnt scan the fd array any more.




2005-02-28 00:06:25

by Paulo Marques

[permalink] [raw]
Subject: Re: Yet another I/O modeling

Quoting Neil <[email protected]>:

> hi,all
> select/poll I/O modeling is fast,easy to use,it returns the number of fd
> which are acitvity,but you should scan the fd array to find which fd is
> ready,it costs a lot of time on a heavy load server.
>
> I proposal another I/O modeling,named eselect.
> struct eselect_struct{
> unsigned long readbitmap[MAXBITMAPFD];//suppose the MAXBITMAPFD is
> MAX OPEN //FD NUMBER PER PROCESS
> int ret;//-1 on error,non-negative return value is the number of
> //acitvity fds
> }
> we can use __set_bit(),__clear_bit(),__find_first_bit()....functions to
> maintain the bitmap.
> So,you shouldnt scan the fd array any more.

I really don't see why __find_first_bit() is better than "scan the fd array".
They both seem like O(N) operations. You might be dividing the tests by 32 (or
64) but in a busy server with 3000 open sockets, what you really want is no
scanning at all.

Anyway, your problem is already solved. You should check epoll:

http://epoll.hackerdojo.com/

It's been in the kernel for a while, now (with CONFIG_EPOLL=y).

--
Paulo Marques - http://www.grupopie.com

All that is necessary for the triumph of evil is that good men do nothing.
Edmund Burke (1729 - 1797)