[svn.haxx.se] · SVN Dev · SVN Users · SVN Org · TSVN Dev · TSVN Users · Subclipse Dev · Subclipse Users · this month's index

Re: 1.7 Performance via HTTP

From: Mark Phippard <markphip_at_gmail.com>
Date: Mon, 16 May 2011 13:52:11 -0400

On Sun, May 15, 2011 at 6:37 AM, Justin Erenkrantz
<justin_at_erenkrantz.com> wrote:
> On Thu, May 12, 2011 at 5:57 PM, Justin Erenkrantz
> <justin_at_erenkrantz.com> wrote:
>> Are you by chance using SSL?
>>
>> I'm seeing something like a 2x perf drop on ra_serf with SSL to an SSL
>> server on the other side of the US from where I am now.  But, over
>> HTTP, ra_serf is pretty close to ra_neon.  (And, the server is 1.6 -
>> not 1.7.)
>
> I've committed a bunch of fixes and improvements to serf for SSL.
> This should bring ra_serf in line with ra_neon for SSL.
>
> As a data point, 'svn ls -v' on a directory - old serf exchanged 90
> TCP packets for 31.5KB; new serf does 41 TCP packets for 11.5KB;
> ra_neon does 99 TCP packets for 19.5KB.  Wall clock time looks to be
> about the same between ra_neon & ra_serf (I'm at high latency).
>
> Basic tests:
> | 1.7.0-dev (serf) | rNNNNNNNN | 1:07.272 | 0:35.990 | 0:04.994 |
> 0:00.043 | 0:00.060 | 0:07.897 | 0:00.041
> | 1.7.0-dev (neon) | rNNNNNNNN | 1:17.145 | 0:23.997 | 0:06.000 |
> 0:00.045 | 0:00.059 | 0:09.893 | 0:00.039

I rebuilt using latest HEAD of Serf and SVN (no patches applied) and
re-ran the tests using a 1.7 serf client and both a 1.6 and 1.7 SVN
server.

While the server was configured according to your recommendations for
KeepAlive, I did not have mod_deflate turned on. Also, I am using
plain HTTP, not SSL.

The numbers were basically the same as I posted before, so I will not
post them again. Your fixes seemed to be related to SSL, so I am
guessing this is not really surprising, but thought I would share
anyway.

One thing I am a little confused about, but maybe it is a question for
C-Mike. When Serf is used, the number of HTTP requests does not go
down very much. 81,938 -> 80,928 I imagine this is because Serf
already did not do all of the PROPFIND nonsense we do with Neon?
Still, what are the HTTPv2 benefits that Serf is supposed to see? I
seem to only see benefits when using Neon.

-- 
Thanks
Mark Phippard
http://markphip.blogspot.com/
Received on 2011-05-16 19:52:40 CEST

This is an archived mail posted to the Subversion Dev mailing list.

This site is subject to the Apache Privacy Policy and the Apache Public Forum Archive Policy.