Hello Luca, It’s a problem with the limited fifo queue size of http2.
Stefan send me a workaround not sure if this one will go into http2 or if he will release a real fix soon. Greets, Stefan Excuse my typo sent from my mobile phone. > Am 02.12.2017 um 09:23 schrieb Luca Toscano <[email protected]>: > > Hi everybody, > > did I miss an update or are we still waiting for more data? (Don't mean to > rush you Stefan, just to understand what's the status of the thread :) > > Luca > > 2017-11-24 15:26 GMT+01:00 Stefan Priebe - Profihost AG > <[email protected]>: >> Thanks i‘ll post a log tonight with a 120s stalled request. >> >> Greets, >> Stefan >> >> Excuse my typo sent from my mobile phone. >> >>> Am 23.11.2017 um 17:09 schrieb Stefan Eissing >>> <[email protected]>: >>> >>> Hey, >>> >>> could you try the patch below and produce such a lovely log file again? >>> H2MaxWorkers please back to before, unconfigured. Thanks! This is a small >>> change that a) logs the interaction with h2_workers a bit more and makes >>> sure that time gets lost where I think it does. It also switches the fifo >>> queue in set mode where duplicate entries are checked, in case that >>> interferes here. >>> >>> Cheers, >>> >>> Stefan >>> >>> <h2worker_register-v0.diff> >>> >>> >>>> Am 23.11.2017 um 14:16 schrieb Stefan Priebe - Profihost AG >>>> <[email protected]>: >>>> >>>> Hi, >>>>> Am 23.11.2017 um 14:10 schrieb Stefan Eissing: >>>>> Interesting. I assume that otherwise this host is the same (OS/CPU etc.) >>>>> as others where it runs without probs? >>>> >>>> Yes and no i got some more reports by colleagues where they've disabled >>>> http2 as the customers had unexpected long loading times. >>>> >>>>> We are not ghosted by some strange blabla-lake hyper threading thingie >>>>> singularity? >>>> >>>> Huhoh what's that? Any chance to add some more debugging? >>>> >>>> Greets, >>>> Stefan >>>> >>>>> >>>>> Need to think about this. >>>>> >>>>>> Am 23.11.2017 um 13:43 schrieb Stefan Priebe - Profihost AG >>>>>> <[email protected]>: >>>>>> >>>>>> *argh*, i was too fast no it did NOT fix the problem. It even happens >>>>>> with: >>>>>> H2MaxWorkers 4096 >>>>>> >>>>>> Sorry about that. >>>>>> >>>>>> Stefan >>>>>> >>>>>>> Am 23.11.2017 um 13:42 schrieb Stefan Priebe - Profihost AG: >>>>>>> Hello,, >>>>>>> >>>>>>> setting: >>>>>>> H2MaxWorkers 1024 >>>>>>> >>>>>>> fixes the issue for me. The main problem is how to i know how many >>>>>>> workers are needed? How can i detect whether all workers of h2 are busy? >>>>>>> >>>>>>> Stefan >>>>>>> >>>>>>>> Am 22.11.2017 um 13:23 schrieb Stefan Priebe - Profihost AG: >>>>>>>> Hell Stefan, >>>>>>>> >>>>>>>> will send a log to you in a few seconds via private email. >>>>>>>> >>>>>>>> Greets, >>>>>>>> Stefan >>>>>>>> >>>>>>>>> Am 21.11.2017 um 23:18 schrieb Stefan Eissing: >>>>>>>>> sorry for the late reply. for stucks trace2 is best. >>>>>>>>> >>>>>>>>>> Am 21.11.2017 um 19:35 schrieb Stefan Priebe - Profihost AG >>>>>>>>>> <[email protected]>: >>>>>>>>>> >>>>>>>>>> Hello Stefan, >>>>>>>>>> >>>>>>>>>> which loglevel do you need? trace2? >>>>>>>>>> >>>>>>>>>> Greets, >>>>>>>>>> Stefan >>>>>>>>>> >>>>>>>>>>> Am 21.11.2017 um 16:48 schrieb Stefan Eissing: >>>>>>>>>>> Never done this, but https://www.howtoforge.com/setenvif_apache2 >>>>>>>>>>> seems like one way to do make it work. >>>>>>>>>>> >>>>>>>>>>>> Am 21.11.2017 um 16:16 schrieb Stefan Priebe - Profihost AG >>>>>>>>>>>> <[email protected]>: >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>>>> Am 21.11.2017 um 16:06 schrieb Stefan Priebe - Profihost AG: >>>>>>>>>>>>>>> Am 21.11.2017 um 15:45 schrieb Stefan Eissing: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Am 21.11.2017 um 14:33 schrieb Stefan Priebe - Profihost AG >>>>>>>>>>>>>>> <[email protected]>: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Hello Stefan, >>>>>>>>>>>>>>> Hello Yann, >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> me the http2 bug tester is calling again ;-) >>>>>>>>>>>>>> >>>>>>>>>>>>>> And the day was going so well... >>>>>>>>>>>>> >>>>>>>>>>>>> I'm sorry ;-) >>>>>>>>>>>>> >>>>>>>>>>>>>>> While running two bash curl while loops the one using http1.1 >>>>>>>>>>>>>>> always >>>>>>>>>>>>>>> finishes in < 0.05s while the http2 one takes sometimes 0.4 to >>>>>>>>>>>>>>> 20s to >>>>>>>>>>>>>>> finish. Sadly i can't reproduce this all the time - mostly more >>>>>>>>>>>>>>> requests >>>>>>>>>>>>>>> more failures. As this is a production server i've no idea how >>>>>>>>>>>>>>> to debug >>>>>>>>>>>>>>> as the http2 trace logs might flood the harddisk. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Hmmm. Do you know if this happens waiting for a response or at >>>>>>>>>>>>>> the end of a connection? Or in the middle of a body? All GETs or >>>>>>>>>>>>>> also POSTs? >>>>>>>>>>>>> >>>>>>>>>>>>> My Test only contains GET - but most probably there are also >>>>>>>>>>>>> running >>>>>>>>>>>>> POST requests but not started by me. >>>>>>>>>>>>> >>>>>>>>>>>>> Strangely this only happens between 1pm and 2pm a day but i've no >>>>>>>>>>>>> idea >>>>>>>>>>>>> what's different at that time. >>>>>>>>>>>> >>>>>>>>>>>> OK i'm also able to reproduce this whenever your want. Can we >>>>>>>>>>>> activate >>>>>>>>>>>> trace logging for a specific IP? So i can generate a http2 log? >>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>>> I can output a lot of information from curl: >>>>>>>>>>>>> time_namelookup >>>>>>>>>>>>> time_connect >>>>>>>>>>>>> time_appconnect >>>>>>>>>>>>> time_pretransfer >>>>>>>>>>>>> time_redirect >>>>>>>>>>>>> time_starttransfer >>>>>>>>>>>>> >>>>>>>>>>>>> Another way might be to enable trace logging only for "my" IP? Is >>>>>>>>>>>>> something like this possible? >>>>>>>>>>>>> >>>>>>>>>>>>> Greets, >>>>>>>>>>>>> Stefan >>>>>>>>>>> >>>>>>>>> >>>>> >>> >
