Hello!

On Thu, Apr 28, 2011 at 08:12:59AM -0400, darckos wrote:

> Hello,
>
> I use nginx 0.8.53 on Solaris 10 and I feel like I reach a kind of limit
> but I'm not abble to find it.
> I use Nginx to serve static image and sometimes to get one image I need
> to wait 10 or more seconds.

Most likely nginx actually disk bound and worker processes are
blocked on disk, and that's what causes such delays.

You may want to examine iostat output to see if it really happens.

> nginx is launched with a ulimit -n 65536, and here is my config:
>
> worker_processes 40;
> error_log logs/error.log;
> events {
> worker_connections 8192;
> devpoll_events 1;
> }
> http {
> include mime.types;
> default_type application/octet-stream;
>
> sendfile on;

[...]

> Do you have any idea to solve this problem ?

Try the following (in no particular order, you may want actually
to start with (3)):

1. Tune worker_processes to add more (or less) concurency to make
sure disks are reasonably loaded but not overloaded (i.e. disk r/w
times are small enough).

2. Try switching accept_mutex off to get better request
distribution between worker processes.

events {
accept_mutex off;
...
}

3. Switch off sendfile and tune output_buffers to use 1-2 big
enough buffers to make sure requests to disks will be big (and
disks aren't trashed with large number of small requests).
Something like

output_buffers 2 256k;

will do the trick (using bigger buffers may help better if you
have enough memory and your images are big enough).

Ideally aio should be used in such situation, but it's not
available with nginx under Solaris.

Maxim Dounin

_______________________________________________
nginx mailing list
nginx@nginx.org
http://nginx.org/mailman/listinfo/nginx