server_names scaling

[Maxim Dounin]

Hello!

On Wed, Apr 16, 2014 at 08:38:10PM +1000, SplitIce wrote:

> Hi all,
> 
> I have spent the day troubleshooting why one server in our network reloaded
> / tested configuration extremely slowly.
> 
> We have found that server_names scales very poorly, once a certain point is
> reached (approx 5.5k entries globally, 5k entries for a single host)
> performance drops from a <0.5s reload time to 15s+.
> 
> The large host of ~5,000 entries is a malware domain zone and all server
> names in this zone are using the wildcard name format.
> 
> For now we have resolved this issue by fixing an inefficiency in our
> configuration (namely using *.domain.com and domain.com) however I feel
> this is most likely a bug or at-least an unintended behaviour.
> 
> Relevant configuration entries:
> server_names_hash_max_size 8000;
> server_names_hash_bucket_size 128;

With max_size 8000, and 5k entries - probability of collisions 
while building a cache is high (think of birthday paradox).  And 
bucket_size 128 isn't high enough to allow multiple collisions.  
As a result, nginx may (and likely will) spend a lot of time trying 
to build an optimal hash.

Trivial solution is to use higher max_size and/or bucket_size.

-- 
Maxim Dounin
http://nginx.org/