Questions about proxy logging and a sanity check

[Igor Sysoev]

On Sat, Apr 19, 2008 at 07:51:39PM -0700, mike wrote:

> I am attempting to replace LVS with nginx for load balancing.
> 
> Basically, one front-end load balancer (running nginx) to 3 backend
> servers (soon to be running nginx + fastcgi/php)
> 
> Using these variables, it looks like it shows -all- the defined
> upstreams it had to try.
> 
> $upstream_addr
> $upstream_response_time
> $upstream_status
> 
> I do realize it orders them differently. Is the first one listed
> always the one that it used? i.e.
> 
> log_format  main  '$upstream_addr - $upstream_response_time - $upstream_status
> 
> I looked to see if there was a variable that had the value of the
> server that replied...
> 
> On a successful request, it shows a simple one line.
> 10.13.5.14:80 - 0.001 - 200
> 
> On a non-successful request (i.e. a 404) I see:
> 10.13.5.12:80, 10.13.5.14:80 - 0.000, 0.001 - 404, 404
> 
> Or, does it try each backend one at a time, and it just happens so
> fast it looks like it's done in parallel? and is it safe to say the
> first one listed is the best option?

nginx tries them sequentially. As to 404, it seemd that you set

     proxy_next_upstream  ...  http_404 ...;

The $upstream_... variables show all tries.

> I plan to use this frontend nginx server for:
> - ssl (if i need it) - client<->frontend nginx (SSL). frontend nginx
> <-> backend (non-SSL)
> - gzipping/compression - client<->frontend nginx (gzip). frontend
> nginx <-> backend (uncompressed)
> - expires headers (perhaps) ..

nginx can do all this.

> I am doing maybe 5 or 6 million requests per day right now. All of
> them would be proxied through this frontend nginx server. nginx will
> be running on a quad-core xeon 3220 w/ 2 gigs of ram available for
> this. Would it be a problem?

SSL and gzip are CPU hogs. I can not say about SSL.

Cluster that runs www.rambler.ru/images.rambler.ru/etc serves about
500-900 millions requests per day. Most of them are static files,
but about 22 millions are proxied and some of them are gzipped.

Cluster has several computers, but 2 computers with Athlon 64 X2 Dual
Core 4200+ / 4G are enough.

When computer handles about 7000 request/s, then network interrupt handler
becames one of the CPU hog, but this is not your case.


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