Re: [Xen-devel] Questions about GPLPV stability tests

[Roderick Colenbrander <thunderbird2k@xxxxxxxxx>]

On Tue, Nov 29, 2011 at 10:15 AM, Andreas Kinzler <ml-xen-devel@xxxxxx> wrote:
>>> Not in this year of my stability tests. In this year I am always
>>> experiencing crashes of domU only. dom0 was always stable.
>>> But last year, I hunted a very serious problem which causes nasty
>>> hangs/crashes in dom0 (which crashes domU as a consequence). See this
>>> mailing list post:
>>> http://lists.xen.org/archives/html/xen-devel/2010-09/msg00556.html
>>> In my tests it clearly shows that if you have a CPU without ARAT and you
>>> don't have the patch from my post, your Xen 4.0.1 or 4.1.1 will crash
>>> under
>>> load and/or after a while. What is your CPU?
>>
>> Most of our machines use i7 950 CPUs. They don't seem to have ARAT.
>
>
> Yes, i7 950 does not have ARAT as it is the first Nehalem generation.
>
>
>> Some other machines use Xeon CPUs with ARAT support. We never had
>> issues on the Xeon systems, so we may actually be suffering from the
>> ARAT issue. Are you still using the patch you linked to in a
>> production environment?
>
>
> Absolutely. As I mentioned I just re-performed tests recently and found that
> even Xen 4.1.1 (earlier tests were for 4.0.1) is unstable without my patch
> on non-ARAT-CPUs.
>
>
>> I wonder why a cleaned up patch like that never made it into core.
>
> The patch actually only disables HPET broadcast which has some downsides
> because it effectively disables C3 states. Without C3 states Nehalem and
> later CPUs cannot enter turbo-mode. So there is some loss of performance.

Would disabling any low CPU power states and turbo clocks in the BIOS,
help as well? Just curious. I have seen other 'weird' performance
issues between machines using the same hardware. Some CPU intensive
algorithm could be twice as slow running on Dom0 compared to the same
kernel without Xen. On other identical systems I didn't see that
issue. I didn't have time to investigate, but I felt there may have
been BIOS setting differences.

>
>
>> I'm going to do some testing (may take a while).
>
>
> Please let me know. There are already some people confirming that HPET
> broadcast is buggy. With more evidence I will contact Keir again and suggest
> to have it (HPET broadcast) fixed or removed.
>
> Regards Andreas

Thanks,
Roderick

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