On 15.09.2013 05:31, Rainer Jung wrote: > On 10.09.2013 16:13, Jim Jagielski wrote: >> For completeness, a full, combined patch is: >> >> http://people.apache.org/~jim/patches/httpd-2.4-event-test.patch >> >> It requires a patch that knows about creating new files >> when encountering /dev/null... > > The code (plus r1410004) runs on eos (US incarnation of www.apache.org) > since a few hours. It runs stably without noticable hickups. > > .4.7-dev-r1523261-skiplist: http://www.us.apache.org/server-status > 2.4.6: http://www.eu.apache.org/server-status > > Currently usr cpu is 60% higher per request, sys cpu is 30% lower, total > cpu is 27% higher. This might be due to differences in the request mix > between the US and EU incarnation, hardware, config etc. are the same. I > plan toswitch the versions between the two and do the comparison again. > > Note that I'm not talking about the CPU percentage value in > server-status which is for the whole system. There the numbers differ a > lot because eos also carries the wiki but that runs outside of the httpd > processes. I'm only measuring the httpd process cpu time per request.
I switched versions between EU and US mirror (aurora and eos) and the differences in CPU are due to the servers and/or request mixes, not due to the new httpd version. You can have a look yourself: Load in requests/s: http://people.apache.org/~rjung/skiplist/load.png usr cpu per request: http://people.apache.org/~rjung/skiplist/usr.png sys cpu per request: http://people.apache.org/~rjung/skiplist/sys.png total cpu per request: http://people.apache.org/~rjung/skiplist/total.png I'm pretty sure from those pictures you would not be able to find the point in time where I switched 2.4.6 and 2.4.7-dev between the servers. It was on 2013-09-15 at 10:36. Here are the pictures showing only one hour around the switch: Load in requests/s: http://people.apache.org/~rjung/skiplist/load_detail.png usr cpu per request: http://people.apache.org/~rjung/skiplist/usr_detail.png sys cpu per request: http://people.apache.org/~rjung/skiplist/sys_detail.png total cpu per request: http://people.apache.org/~rjung/skiplist/total_detail.png Regards, Rainer >> On Sep 10, 2013, at 9:52 AM, Jim Jagielski <j...@jagunet.com> wrote: >> >>> For the testing, we need: >>> >>> http://people.apache.org/~jim/patches/httpd-2.4-skiplist.patch >>> http://people.apache.org/~jim/patches/httpd-2.4-podx.patch >>> http://people.apache.org/~jim/patches/httpd-2.4-event.patch >>> >>> This includes all the performance/sync updates >>> >>> Can we get infra to test that out somewhere? >>> >>> On Sep 6, 2013, at 7:19 AM, Jim Jagielski <j...@jagunet.com> wrote: >>> >>>> >>>> On Sep 5, 2013, at 5:46 PM, Rainer Jung <rainer.j...@kippdata.de> wrote: >>>> >>>>> On 05.09.2013 16:35, Jim Jagielski wrote: >>>>>> BTW, the main diff between event in trunk and 2.4 is >>>>>> the use of skiplist. My benchmarks show decreased latency >>>>>> and a performance boost of ~5% (on avg). Can anyone confirm? >>>>>> It would be nice to possibly get that in 2.4.7 as well. >>>>> >>>>> I can offer to negotiate with infra to run 2.4 head plus event patches >>>>> for skiplist use on www.a.o EU if that helps collecting votes. >>>>> >>>>> I did the httpd updates on www.a.o during the last months and in >>>>> previous years running non-released stuff there was a common practice, >>>>> so I hope infra will trust our opinion on whether the code should be >>>>> ready for prime time or not. >>>>> >>>>> If we want to see the code running there, I would appreciate a patch >>>>> ready to apply to 2.4 event.c. I currently feel not in the position to >>>>> judge, whether running that code on www.a.o actually is a good idea or >>>>> not. Opinions? >>>> >>>> >>>> Great idea. I can create the 2.4 patch. (and add it to STATUS) ;)
