Date: Wed, 15 Aug 2012 14:07:32 +0300 From: Alexander Motin <mav@FreeBSD.org> To: lev@FreeBSD.org Cc: Adrian Chadd <adrian@freebsd.org>, Doug Barton <dougb@FreeBSD.org>, current@freebsd.org Subject: Re: CURRENT as gateway on not-so-fast hardware: where is a bottlneck? Message-ID: <502B82F4.1090804@FreeBSD.org> In-Reply-To: <1849591745.20120815144006@serebryakov.spb.ru> References: <157941699.20120815004542@serebryakov.spb.ru> <CAJ-Vmon86-FPs4%2BXXkQXAow1jW465pMM2Sj7ZHi_0_E9VYSFSA@mail.gmail.com> <502AE8B5.9090106@FreeBSD.org> <502B775D.7000101@FreeBSD.org> <1849591745.20120815144006@serebryakov.spb.ru>
next in thread | previous in thread | raw e-mail | index | archive | help
On 15.08.2012 13:40, Lev Serebryakov wrote: > You wrote 15 августа 2012 г., 14:18:05: > AM> It is quite pointless to speculate without real info like mentioned > AM> above KTR_SCHED traces. Main thing I've learned about schedulers, things > AM> there never work as you expect. There are two many factors are relations > AM> to predict behavior in every case. > I'll take these with as much variants (ULE and 4BSD, polling with > HZ=1000 and interrupts with default HZ) as I can, in day or two. > Now I have kernels with KTR compiled in (GEN, NET and SCHED). > > AM> About Soekris and idle CPU measurement, let's start from what kind of > AM> eventtimer is used there. As soon as it is UP machine, I guess it uses > AM> i8254 timer in periodic mode. It means that it by definition can't > It doesn't have any other timers. You could think about this machine > as about good old "true" i386, with PCI (and some additional fancy > commands in CPU core, something like classic Pentium) but > nothing more. > > kern.eventtimer.choice: i8254(100) RTC(0) > kern.eventtimer.et.RTC.flags: 17 > kern.eventtimer.et.RTC.frequency: 32768 > kern.eventtimer.et.RTC.quality: 0 > kern.eventtimer.et.i8254.flags: 1 > kern.eventtimer.et.i8254.frequency: 1193182 > kern.eventtimer.et.i8254.quality: 100 > kern.eventtimer.periodic: 1 > kern.eventtimer.timer: i8254 > kern.eventtimer.activetick: 1 > kern.eventtimer.idletick: 0 > kern.eventtimer.singlemul: 2 Yes, that is what I expected to see there. If you have timecounter other then i8254, you can release i8254 from those duties to allow using it as one-shot setting hint.attimer.0.timecounter=0. Otherwise there are no options now. > AM> properly measure load from treads running from hardclock, such as > AM> dummynet, polling netisr threads, etc. > You see, here are two different problems: > > (a) with polling, system is responsive under any load, but wire2wifi > performance is hugely affected by wire2wire traffic (and mpd5 > inbetween). And, yes, "top" seems to lie about idle time. I don't know why wifi is so different. Suppose it is for some reason more affected by latencies. > (b) with interrupts, system works much better when it works (wire2wifi > speed is affected by wire2wire traffic, but to much less extent), but > it freezes every third minute for minute, when traffic is passed, but > no user-level applications including BIND and DHCP server) works at > all FOR MINUTE OR MORE. It not looks like 100ms lag, which could affect > video playback. It looks like 60-120 seconds lag! At least, in case of > ULE, I didn't try 4BSD yet. In this case problem may be that kernel and interrupt threads are all having absolute priorities. It means until they release the CPU, user-level may get no CPU time at all. :( -- Alexander Motin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?502B82F4.1090804>