On Wed, Apr 22, 2009 at 09:56:37AM +0800, Delta Yeh wrote:

> 2009/4/22 Igor Sysoev <is@rambler-co.ru>
>
> > On Tue, Apr 21, 2009 at 04:43:05PM +0200, Roger Pack wrote:
> >
> > > > Nginx server in front could take care of. I'd rather see something like
> > > > Libevent (i.e. an asynchronous HTTP server engine) with embedded
> > Python.
> > > > The core of Nginx is probably similar to Libevent with its HTTP layer.
> > > > Would be interesting to find out how Nginx core compares to Libevent
> > > > with HTTP layer in terms of performance.
> > >
> > > Being somewhat of a newbie, just throwing this out.
> > > It would indeed be interesting to see if nginx with a libev core would
> > > be faster than the current (ex:
> > >
> > http://www.zenebo.com/word/asynchronous-programming/lighttz-a-simple-and-fast-web-server/
> > )
> > >
> > > Though probably wouldn't be tons faster.
> > > Thoughts?
> >
> > I do not think that using libev in nginx will change anything.
> >
> > Probably, the following settings may slightly improve or worsen:
> >
> > 1) turing sendfile off: it may be not effective on small files.
> > Also not that lighttz does not read file at all each request.
> >
> > 2) using open file cache, it saves 3 syscalls per request (open, fstat,
> > close):
> >
> > open_file_cache max=1000 inactive=20s;
> > open_file_cache_valid 30s;
> > open_file_cache_min_uses 2;
> > open_file_cache_errors on;
> >
> > 3) using buffered log:
> >
> > access_log /path/to/access.log buffer=32k;
> >
> > or turing it off at all:
> >
> > access_log off;
> >
> > 4) using 1 worker.
> >
>
> 1 worker for each core?

No, just 1 worker. In the test there is already 1 worker per core (dual core).
In microbenchmark (not real life) this may improve or worsen.


--
Igor Sysoev
http://sysoev.ru/en/