Re: PIT, TSC and power management [was: Re: 2.6.0-test3 "loosingticks"]

[Charles Lepple]

john stultz said:
> On Thu, 2003-08-14 at 17:19, Charles Lepple wrote:
>> I also see the time offset problem (Athlon MP 2000+ x2, Tyan S2460 m/b,
>> 2.6.0-test{1,2,3}) but it is most noticeable when I have amd76x_pm
>> installed (it's not in 2.6.x yet, but a late 2.5.x patch was posted to
>> LKML a little while back).
>>
>> amd76x_pm is roughly equivalent to ACPI C2 idling, but since my BIOS
>> doesn't export any C-state functionality to the kernel ACPI code, I am
>> stuck with letting amd76x_pm frob the chipset registers. A quick look at
>> AMD's datasheets does not indicate that a return from C2 should cause
>> much delay at all-- if I understand the timing requirements correctly,
>> it would have to sit for more than 1 ms to miss more than one interrupt.
>> That said, I don't see any missing interrupts indicated in
>> /proc/interrupts, nor do any such messages appear in the kernel logs.
>
> In this case you're throttling the cpu frequency. This affects the
> frequency the TSC updates, which makes it very hard to use the TSC as a
> timesource (the cpu_freq notifier tries to compensate by changing the
> tsc multiplier but my systems don't have cpu_freq drivers, so I've not
> seen it work).

I'm not familiar with the cpu_freq code, or how true ACPI throttling is
implemented, but it sounds like the amd76x_pm driver is doing something a
little different than throttling. I tried the regular ACPI code on an IBM
desktop, and its throttling support appears to offer several distinct
throttling percentages, which would seem to be much easier to compensate
for. The amd76x_pm idle routine simply sets a bit in one of the bridge
chips, but I think that it turns the clocks back on at some indeterminate
time in the future (probably triggered by an interrupt) which would be
hard to measure if clocks are stopped.

>> Brings up another question: does the "try HZ=100" suggestion still apply
>> for these faster machines? I would think that if HZ=1000 is too fast,
>> then at least an occasional lost interrupt would be logged.
>
> If you're losing interrupts and the lost-tick detection code is not
> compensating, shifting back to HZ=100 just tries to minimize the
> problem.

OK. Well, I'm optimistic, so I'll try that and see if that pulls the error
down to a point where ntpd can manage to control the clock.

>> When using the TSC for time-of-day, I generally have to set tick to
>> 10200 or somewhere thereabouts. ntpd usually gives up after a few hours,
>> though, so I presume that this value for tick is only good for a certain
>> combination of processor load and planetary alignment.
>>
>> I booted with clock=pit to test that, and now I need tick=9963
>> (according to adjtimex's configuration routine). However, that makes the
>> clock jump all over the place, with ntpd making step adjustments +/- 2
>> seconds every 5 minutes.
>>
>> > Approximately at what rate does it skew?
>>
>> Well, it's not constant, and I don't trust the tick values given above,
>> since they don't seem to hold true for long.
>
> Do these problems still show when you're not using the amd76x_pm?

I'll have to try again next week-- I think the last time that I tried to
remove amd76x_pm, I didn't reset the ntp adjustments (drift, plus the
adjtimex variables) so it was fighting the old power-management values.

>> > Does ntpdate -b <server></server> set it properly?
>>
>> I'm confused. Are there cases where a step time adjustment would fail?
>> Is there a possibility that the kernel is rejecting ntpd's step
>> adjustments? (I presume that these use the same as 'ntpdate -b';
>> specifically, the time is not slewed.)
>
> Well, depending on how ntp is compiled, it could use stime, rather then
> settimeofday. This causes ntp to set the time on average .5 seconds off
> the desired time. Since .5 is outside the .128 sec slew boundary, ntp
> will do another step adjustment which has the same poor accuracy. This
> results in ntp just hopping back and forth around the desired time.

I'll have to check with strace (I'm using ntpd from Debian sid).

thanks,

-- 
Charles Lepple <clepple@xxxxxx>
http://www.ghz.cc/charles/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/