On Tue, Jun 21, 2011 at 7:36 AM, Stefan Sperling <stsp_at_elego.de> wrote:
> [ Moving to dev@ ]
>
> On Tue, Jun 21, 2011 at 02:27:55PM +0200, OBones wrote:
>> Uwe Schuster wrote:
>> >Stefan Sperling wrote:
>> >>On Tue, Jun 21, 2011 at 02:02:08AM +0200, Uwe Schuster wrote:
>> >>>Stefan Sperling wrote:
>> >>>
>> >>>>>1. Performance
>> >>>>
>> >>>>Can you set
>> >>>> http-library=neon
>> >>>>in the '[global]' section of your 'servers' configuration file
>> >>>>and see if that helps?
>> >>>>
>> >>>>The checkout finishes within 2minutes for me (with neon, on OpenBSD,
>> >>>>16mbit/s downstream).
>> >>>
>> >>>Yes this helps very much!
>> >>>
>> >>>39 seconds and 6,82 MB instead of 30 min 27 s and 55,70 MB
>> >>
>> >>Right. So this is an issue with serf.
>> >>None of the existing serf issues seem to match, though.
>> >>
>> >>Is there anything special about your setup, such as connecting through a
>> >>proxy?
>> >
>> >No, there is no proxy involved and there shouldn't be anything
>> >special with my setup. On the PC a personal firewall (Comodo) is
>> >running and the PC is connected over a Gigabit switch to a T-Com
>> >router with DSL modem (SP 700). I did run the tests with TSVN 1.6
>> >an 1.7 on the same machine (they are installed in parallel).
>> >
>> Well, you don't have a proxy installed, but connecting to a
>> SourceForge hosted server is guaranteed to go through a proxy
>> because they use it as a means of load balancing. It's transparent
>> to the end user, but maybe not to the libraries
>
> It would be nice to get an official confirmation about this from SF.net.
>
> Because if this is the case, I think we should seriously reconsider the
> decision of making serf the default for 1.7. We definitely cannot afford
> to break Subversion for all sourceforge users in such a fundamental way.
> People who can benefit from serf's performance characteristics without
> facing such problems can still enable it.
One other data point: A few months ago the ghost of Ben
Collins-Sussman showed up on IRC asking about ra_serf and 1.7. He
made the observation that from a Google hosting perspective, the flood
of GET requests from serf was about the same as a botnet attempting to
DoS the server. Neither he nor I (nor anybody else in IRC at the
time) was able to come up with a good way of differentiating the two.
I've copied Ben on this mail, so that perhaps he can provide an
update.
It could be possible that SF.net is flagging the multitude of GET
requests as a DoS and intentionally throttling them. But that's just
a guess.
-Hyrum
Received on 2011-06-21 15:16:46 CEST