You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@subversion.apache.org by Justin Erenkrantz <ju...@erenkrantz.com> on 2011/01/23 18:27:18 UTC
Tuning recommendations for Apache HTTP Server w/ra_serf clients (MaxKeepAliveRequests)
I think this probably belongs in the book, but I don't know exactly
where...so I'll ensure that it's noted again and let someone else
figure out where it should live.
As we start to have more users with ra_serf, we should also recommend
that httpd server admins bump up their MaxKeepAliveRequests from 100
(default) to at least 1000. This'll help ra_serf utilize TCP
connections more efficiently - as ra_serf issues 2 HTTP requests per
file that it needs to check out or update.
ra_serf has heuristics so that it'll work gracefully under the default
settings (and determine what the value actually is), but you'll get
better network performance if you can utilize fewer TCP connections.
This way, ra_serf will only need to open a new connection every 500
files rather than every 50. (Honestly, there is really no reason that
it couldn't be 10000.)
If we also have a tuning section, we should remind folks to also
enable mod_deflate (SetOutputFilter DEFLATE in their Location block)
as that'll help a bit when transferring XML content back and forth.
mod_deflate will come at a slight latency penalty, but that'll be
offset for folks with slower connections taking less time to transfer
the responses overall.
FWIW, I've already asked ASF infra to bump up the MaxKeepAliveRequests
settings for svn.apache.org.
Thanks! -- justin
Re: Tuning recommendations for Apache HTTP Server w/ra_serf clients (MaxKeepAliveRequests)
Posted by Justin Erenkrantz <ju...@erenkrantz.com>.
On Sun, Jan 23, 2011 at 11:26 AM, Stefan Sperling <st...@elego.de> wrote:
> On Sun, Jan 23, 2011 at 09:27:18AM -0800, Justin Erenkrantz wrote:
>> If we also have a tuning section, we should remind folks to also
>> enable mod_deflate (SetOutputFilter DEFLATE in their Location block)
>> as that'll help a bit when transferring XML content back and forth.
>> mod_deflate will come at a slight latency penalty, but that'll be
>> offset for folks with slower connections taking less time to transfer
>> the responses overall.
>
> Has this problem been fixed, then?
> http://svn.haxx.se/dev/archive-2009-08/0274.shtml
> I don't think we should recommend mod_deflate if that problem still exists.
No, it doesn't look like it has been fixed yet.
However, how many real-world HTTP clients are out there that don't
speak zlib? =)
> I don't think that thread came to a conclusion.
> Greg was implying the bug was in mod_deflate rather than svn:
> http://svn.haxx.se/dev/archive-2009-08/0290.shtml
Greg is correct in that the patch doesn't go far enough - mod_dav_svn
should change all uses of output to be a ** rather than *.
As far as mod_deflate being broken, yah, it probably shouldn't be
trying to do any memory usage until it knows it is
activated...but...that's a topic for another list (dev@httpd).
mod_deflate should probably memorize the fact that the checks have
already run and get out of the way...probably with setting no-gzip in
subprocess_env? -- justin
Re: Tuning recommendations for Apache HTTP Server w/ra_serf clients (MaxKeepAliveRequests)
Posted by Justin Erenkrantz <ju...@erenkrantz.com>.
On Sun, Jan 23, 2011 at 11:26 AM, Stefan Sperling <st...@elego.de> wrote:
> On Sun, Jan 23, 2011 at 09:27:18AM -0800, Justin Erenkrantz wrote:
>> If we also have a tuning section, we should remind folks to also
>> enable mod_deflate (SetOutputFilter DEFLATE in their Location block)
>> as that'll help a bit when transferring XML content back and forth.
>> mod_deflate will come at a slight latency penalty, but that'll be
>> offset for folks with slower connections taking less time to transfer
>> the responses overall.
>
> Has this problem been fixed, then?
> http://svn.haxx.se/dev/archive-2009-08/0274.shtml
> I don't think we should recommend mod_deflate if that problem still exists.
No, it doesn't look like it has been fixed yet.
However, how many real-world HTTP clients are out there that don't
speak zlib? =)
> I don't think that thread came to a conclusion.
> Greg was implying the bug was in mod_deflate rather than svn:
> http://svn.haxx.se/dev/archive-2009-08/0290.shtml
Greg is correct in that the patch doesn't go far enough - mod_dav_svn
should change all uses of output to be a ** rather than *.
As far as mod_deflate being broken, yah, it probably shouldn't be
trying to do any memory usage until it knows it is
activated...but...that's a topic for another list (dev@httpd).
mod_deflate should probably memorize the fact that the checks have
already run and get out of the way...probably with setting no-gzip in
subprocess_env? -- justin
Re: Tuning recommendations for Apache HTTP Server w/ra_serf clients
(MaxKeepAliveRequests)
Posted by Stefan Sperling <st...@elego.de>.
On Sun, Jan 23, 2011 at 09:27:18AM -0800, Justin Erenkrantz wrote:
> If we also have a tuning section, we should remind folks to also
> enable mod_deflate (SetOutputFilter DEFLATE in their Location block)
> as that'll help a bit when transferring XML content back and forth.
> mod_deflate will come at a slight latency penalty, but that'll be
> offset for folks with slower connections taking less time to transfer
> the responses overall.
Has this problem been fixed, then?
http://svn.haxx.se/dev/archive-2009-08/0274.shtml
I don't think we should recommend mod_deflate if that problem still exists.
I don't think that thread came to a conclusion.
Greg was implying the bug was in mod_deflate rather than svn:
http://svn.haxx.se/dev/archive-2009-08/0290.shtml
Stefan