Re: getting oprofile out of timer mode

[William C. <wc...@re...>]

Andrei Lifchits wrote:
> Hi Will,
> 
> Can I ask you another question then. We are actually trying to profile
> XenLinux (2.6.16-xen) using Xenoprof, which uses Oprofile (we tried
> the regular Linux 2.6.11 with FC4 just to see if Oprofile would work
> there). We have the same problem with both kernels (i.e. Oprofile
> can't get out of the Timer Mode), but the XenLinux kernel does have
> SMP support. So could the problem be somewhere else in the case of
> XenLinux? (or is this question more appropriate for the Xenoprof
> authors?)
> 
> Thanks,
> Andrei

Did the regular SMP FC4 kernel work? I would suggest first trying the 
regular SMP kernel to make sure that the hardware has the appropriate 
support and that gives access to the performance monitoring hardware in 
the processor. On some pentium III laptops the apic stuff doesn't seem 
to work.

I haven't played around with the xen stuff much lately. Maybe the 
xenoprof authors can give you some ideas what to look for.

IIRC the Red Hat developer have pretty much dropped trying to maintain 
xen for FC4. There might be some hidden problems in there somewhere, but 
it is unlikely to be fixed.

-Will

> 
> 
> On 4/17/06, William Cohen <wc...@re...> wrote:
> 
>>Andrei Lifchits wrote:
>>
>>>Hello,
>>>
>>>We are trying to use the Oprofile tool (for a class project), and we
>>>have the following problem: if we try to specify events to profile
>>>(e.g. --event=DATA_MEM_REFS), the program says it's in the Timer Mode.
>>>The Oprofile FAQ tells us to add the "lapic" boot option in this case
>>>(we use Pentium III, which should have hardware counters), but doing
>>>so does not change anything for us. Oprofile still starts in the Timer
>>>Mode. Could there be any other reasons for this?
>>>
>>>The corresponding line in our grub config file looks like this:
>>>kernel /vmlinuz-2.6.11-1.1369_FC4 ro root=LABEL=/ rhgb quiet selinux=0 lapic
>>>
>>>Thanks,
>>>Andrei
>>
>>Use the SMP kernel instead. It has the needed support for APIC that is
>>needed to route the performance counter overflows to the interrupt hardware.
>>
>>-Will
>>