subversion-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ben Reser <...@reser.org>
Subject Re: SVNCacheRevProps and other performance tweaks.
Date Wed, 14 May 2014 19:06:58 GMT
On 5/14/14, 10:11 AM, Terry Dooher wrote:
> SVNPathAuthz – Currently ‘On’. To my understanding, the value ‘short_circuit’
> will prevent any other path-based authorisation  being used for requests. I’m
> running Apache2.4/mod_dav_svn, with local Basic Auth and local Authz files. No
> network drives or other auth methods are configured for the server, which is
> dedicated to svn. Would short_circuit save me any processing time in this
> scenario or is it only useful for people with LDAP or other auth systems that
> might slow down request processing?

This setting controls the handling of authorization for what I refer to as
secondary paths.  I.E. paths that are accessed other than the path that gets
directly accessed as the uri in the request line.  A good example of this a
'svn log' request will check authorization for each file modified by a revision
when returning the log data for that revision in order to determine how much
information to return (some or all of the revision properties are omitted based
on this).

on (default): This causes the authorization check to generate a sub-request.
This means that a GET request for each of the secondary paths described above
is generated.  This request is re-run through all of the authentication and
authorization processing.  Usually this setting (despite being the default) is
not what you want.  It is however the safest setting from a security
perspective.  Unless you are using a httpd authorization module that also looks
at paths this added overhead does nothing for you.  LDAP and other external
authentication makes the performance of this worse, however it's still a
measurable performance delay without such systems.

off: The secondary path setting does not run.  It presumes if you have access
to the URI in the request-line you have access to all paths.  This is the
setting you want if you are not using mod_authz_svn (i.e. not using path based
authorization).

short_circuit: The secondary path checks do not generate sub-requests but
rather short circuit the httpd authentication/authorization infrastructure and
directly as mod_authz_svn.  This has significantly less overhead, while being
as secure as on provided that the only authorization module that is path aware
is mod_authz_svn.  If you're using authz unless you have an extremely rare case
you want this setting.

In your scenario you want to use short_circuit.

> SVNCacheRevProps – Currently ‘Off’.  I’m told that it “should only be enabled
> under the conditions described in the documentation”, but I can’t find any
> reference to it in the source or the 1.8 book, yet. What are the conditions in
> which it should be used and what happens if I use it outside those conditions?
> Do I risk DB corruption or is it only of benefit in specific scenarios?

Actually I think it's more that you really must have this enabled if you have
revprop packing turned on.  I can't think of a reason you don't want this turn
off other than desiring to reduce resource usage at the cost of slower performance.

The documentation in question is the release notes:
http://subversion.apache.org/docs/release-notes/1.8.html#revprop-caching
http://subversion.apache.org/docs/release-notes/1.8.html#revprop-packing

> SVNInMemoryCacheSize – Just want to confirm this one: Is the overall server
> memory footprint (CacheSize*No. of worker processes) under mod_dav_svn? I’m
> using the prefork mpm (some minor PHP admin scripts on the box), and typically
> have 20-80 running child workers active. Bumping this up 256kb or 1MB would
> have a negligible effect, if this is the case.

First of all you probably want to read the 1.7 release notes on performance tuning:
http://subversion.apache.org/docs/release-notes/1.7.html#server-performance-tuning

Yes the amount of memory footprint of the cache (there will still be more
memory used for other uses by the server) will be CacheSize*No. of worker
process.  Since the memory caches are per-process we do not recommend using
prefork with Subversion, but rather suggest that you use worker or event MPMs.

With the worker or event MPMs you are able to share the cache between the
threads within the same processes.  This allows you to set a larger cache size
since you'll have fewer processes and more of the threads will gain access to
the cache, making it to be more effective.

I'd suggest you run the PHP in a separate instance of httpd or on a separate
machine.  It's going to be difficult to tune httpd for Subversion with these
other uses in the same httpd processes.

Mime
View raw message