Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 21 Aug 2010 14:00:01 +0000
From:      "b. f." <bf1783@googlemail.com>
To:        Andriy Gapon <avg@icyb.net.ua>
Cc:        freebsd-current@freebsd.org
Subject:   Re: Latest intr problems
Message-ID:  <AANLkTinL6TUXHA-w3tgmBnrqyPxPL=2Z8XJ6fsGvonPs@mail.gmail.com>
In-Reply-To: <4C6FD562.1030700@icyb.net.ua>
References:  <AANLkTin0SHWRobn%2BGREOPtyRMGhoLKdFBMXQWMSa%2Brt7@mail.gmail.com> <4C6FD562.1030700@icyb.net.ua>

next in thread | previous in thread | raw e-mail | index | archive | help
On 8/21/10, Andriy Gapon <avg@icyb.net.ua> wrote:
> on 21/08/2010 16:04 b. f. said the following:
>> Andriy Gapon wrote:
>>> on 21/08/2010 12:35 Andriy Gapon said the following:

> Well, let's try to not muddy the waters prematurely.

It's not premature to say that his machine has some peculiar
clock-related features, that should be kept in mind while testing.
This came up earlier:

http://lists.freebsd.org/pipermail/freebsd-current/2010-June/018179.html

and is partly why I recommended some of the settings below.

>
>> Alexander and I recommended
>> that he try different clocks, and just recently, for example, he wrote
>> that he had used:
>>
>> loader.conf
>> hint.apic.0.clock="0"
>> hint.atrtc.0.clock="0"
>> hint.attimer.0.clock="0"
>> hint.hpet.0.legacy_route="1"
>
> Well, I don't see much point in doing the above in this situation.

We suspected clock problems as well, and were trying to isolate the
problem taking other clocks completely out of consideration.  Probably
the legacy_route could have been discarded, but in that case the
second eventtimer  would have to be emulated with NONE.

>
>> machdep.disable_rtc_set="1"
>> kern.eventtimer.timer2="HPET"
>> kern.eventtimer.timer1="NONE" (Or, if available, HPET1, ...)
>
> So, what was actually used here?
> I don't think that NONE is a good idea.

Doug will have to answer that -- he wasn't specific in his reply.

>> kern.eventtimer.singlemul="1"
>>
>> sysctl.conf:
>> kern.timecounter.hardware=HPET
>>
>> and reported that it did not help.  The HPET doesn't usually suffer
>> from the problem that you are describing, right?
>
> Right.
> Still I would prefer that Doug would do the cleaner experiment(s) that I
> suggested.  And if the problem persists then elimination of LAPIC timer
> would
> make the picture clearer (for me).

Sure, let's see if restraining the C-states yields any results.

>
> P.S.
> I still think that KTR+schedgraph would be the best tool here.

Attilio recommended that originally, then (reportedly) changed his
mind and said to try dtrace.  I brought it up again earlier, and Doug
said that he would make some experiments.

b.



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?AANLkTinL6TUXHA-w3tgmBnrqyPxPL=2Z8XJ6fsGvonPs>