httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Brian Pane <brian.p...@cnet.com>
Subject Re: cvs commit: httpd-2.0/docs/manual/mod mod_log_config.html
Date Thu, 10 Jan 2002 04:35:31 GMT
Joshua Slive wrote:

>>From: Brian Pane [mailto:brian.pane@cnet.com]
>>
>
>>>Can you really benchmark speed differences between the two?
>>>
>>Yes.  And the difference is quite large.
>>
>
>You've tried out Apache with the two different methods of exclusion and
>found that the LogExclude method results in a server that can serve
>significantly more traffic?  I'm surprised!
>

My methodology was to extract the cost of the regex and hash lookup
operations from the profile data from the current httpd and compute
the relative cost of the two mechanisms based on those (my last message
describes the data in a bit more detail).

>>I'm mildly alarmed that we're actually advocating it as a general-purpose
>>solution in the documentation URLs listed above; we really should warn
>>readers that the processing cost scales linearly with the number of rules.
>>
>
>There are many directives in apache that work like this, including
>name-virtual hosts, Redirects, Aliases, etc.
>

If I remember correctly, the name-based virtual hosts have been fixed so
that they aren't O(N).  Redirects and Aliases are O(N), though, which is
why we don't use them too liberally on high-traffic servers.

>  My guess is that for 99% of
>cases, all these things put together are lost in the backround noise of
>disk/network speed.
>

Depends on what metric you're looking at.  If you look at delivery speed as
measured on the client, the network tends to hide the impact of CPU-bound
inefficiencies.  But from a data center cost perspective, a much more
important metric is how much CPU time you're burning per request.  An n%
increase in CPU utilization per request translates pretty linearly into
increased hardware cost.

--Brian



Mime
View raw message