Re: [Xen-devel] Re: [Xen-users] ntpd under Xen Dom0 exhibits extremely high jitter/noise? runs stable/quiet under non-xen kernel.

[mail ignored <0.bugs.only.0@xxxxxxxxx>]

hi

On Tue, Jan 19, 2010 at 1:24 AM, Keir Fraser <keir.fraser@xxxxxxxxxxxxx> wrote:
> On 19/01/2010 02:12, "mail ignored" <0.bugs.only.0@xxxxxxxxx> wrote:
>
>> (1) *is* there, in fact, a bug here?
>
> It's not clear. It could alternatively be a problem with the HPET on that
> specific system. Have you reproduced your issues on a different type of
> system (making sure that HPET platform timer is detected and used on that
> system too, of course)? That would then obviously point more at it being a
> software bug. What is the make/model of the system exhibiting the problem,
> by the way?

the 5 boxes on which this is currently tested/reproducible are

mobo + cpu:
(1)(2) ASUS M4A78T-E           AM3 AMD 790GX ATX, 2 @ AMD Phenom II X4 945 cpu
(3)    MSI 790FX-GD70          AM3 AMD 790FX ATX, 1 @ AMD Phenom II X4 920 cpu
(4)(5) GIGABYTE GA-MA770T-UD3P AM3 AMD 770   ATX, 1 @ AMD Phenom II X4 945 cpu
                                                  1 @ AMD Phenom II X4 920 cpu


all opensuse 11.2 ...

>> (2) which Timer is, in keeping with any/all issues discussed @ the
>> aforementioned wiki page, recommended?
>
> HPET or ACPI-PM timers ought to work equally well, assuming bugs haven't
> crept into Xen's code paths for either one, and both are implemented
> correctly by the hardware Xen is running on. I have to say I think it's
> unlikely to be a software bug, because Xen's code paths to use HPET and
> ACPI-PM timers are really very simple indeed -- it's hard to see where a bug
> could exist in there. But of course that doesn't make it absolutely
> impossible.

from reading the wiki page, i got the sense that HPET is "by far"
recommended -- if only due to lack of specific mention of acpi(_pm) as
clocksource.

can't argue with the fact that it works, though.  if "ACPI-PM timers
ought to work equally well", then I've a clear option/solution -- and
knowledge enuf, now, to know where to look for related problems should
they creep up in the future, perhaps on new/different hardware.

>> (3) is the *real* answer simply intended to be "try them all, pick what
>> works"?
>
> Well, really there oughtn't to be enough of a difference between them for
> anyone to really notice or care which is actually being used.

ok.  though, reproducibly across all the aforementioned systems, there is.

> If this does
> turn out to be a hardware issue, then yes that is just the kind of case
> where you might expect to have to tweak things on the Xen command line.

the 'common' hardware is, of course, the AMD Phenom II X4 cpu ...

> One option for us is to change the priority ordering of platform timers to
> prefer ACPI-PM above HPET, when both are available. That would assume that
> ACPI-PM is more reliable than HPET across a wide range of systems though. We
> don't really have much evidence one way or the other.

that really only makes sense, doesn't it, if it's known to be NOT
problematic on other platforms.   given my obviously limited expertise
in the xen source, this all 'smells' like a problem ...

> Another would be to dynamically detect platform timers that aren't working
> well and disable them at run time. That would require some more thought, and
> isn't going to be something for Xen 4.0 at this point.

understood.

pls let me know if there's additional info you need; or, if this
should issue should simply be allowed to 'fade' now that there's a
clear WORKSFORSOME solution.

thanks.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel