perf top broken on ppc64

perf top broken on ppc64

[Alexander Graf]

Hi,

While trying to find out performance bottlenecks in KVM for PowerPC I =
figured I'd try and use "perf top" to see what's going on in the system. =
This works great on my G4, but doesn't on the Powerstation (970MP).

The only weird thing I can imagine about this setup is that I'm running =
32 bit userland on a 64 bit kernel. So I went ahead and compiled perf =
for ppc64 - without any change:

# file `which perf`
/root/bin/perf: ELF 64-bit MSB executable, 64-bit PowerPC or cisco 7500, =
version 1 (SYSV), for GNU/Linux 2.6.4, dynamically linked (uses shared =
libs), not stripped

Whenever I call 'perf top' I get the following output:


=
--------------------------------------------------------------------------=
-----------------------------
   PerfTop:   57159 irqs/sec  kernel:100.0% [1000Hz cycles],  (all, 4 =
CPUs)
=
--------------------------------------------------------------------------=
-----------------------------

             samples  pcnt  DSO
             _______ _____ =20


Where there would usually be functions being listed I don't get =
anything.
The interesting part is that 'perf record -g' and 'perf report -g' work =
just fine. I get reasonable output. I also had 'perf top' with the exact =
same binaries working by accident too once. But then I rebooted and not =
it's all moot again.

So the question is: why doesn't the top function work for me?


Thanks,

Alex

Re: perf top broken on ppc64

[Ian Munsie]

Excerpts from Alexander Graf's message of Wed Apr 21 09:21:36 +1000 2010:
> Hi,
> 
> While trying to find out performance bottlenecks in KVM for PowerPC I
> figured I'd try and use "perf top" to see what's going on in the
> system. This works great on my G4, but doesn't on the Powerstation
> (970MP).
> 
> The only weird thing I can imagine about this setup is that I'm
> running 32 bit userland on a 64 bit kernel. So I went ahead and
> compiled perf for ppc64 - without any change:

I'm using 32 bit userland and 64 bit kernel on a PowerPC box and it's working for me.
Are you building perf from the tip tree?

Cheers,
-Ian

Re: perf top broken on ppc64

[Alexander Graf]

On 21.04.2010, at 07:29, Ian Munsie wrote:

> Excerpts from Alexander Graf's message of Wed Apr 21 09:21:36 +1000 =
2010:
>> Hi,
>>=20
>> While trying to find out performance bottlenecks in KVM for PowerPC I
>> figured I'd try and use "perf top" to see what's going on in the
>> system. This works great on my G4, but doesn't on the Powerstation
>> (970MP).
>>=20
>> The only weird thing I can imagine about this setup is that I'm
>> running 32 bit userland on a 64 bit kernel. So I went ahead and
>> compiled perf for ppc64 - without any change:
>=20
> I'm using 32 bit userland and 64 bit kernel on a PowerPC box and it's =
working for me.
> Are you building perf from the tip tree?

I'm using kvm.git which is pretty close to tip. The version says =
something 2.6.34-rc3'ish. Has anything significantly changed since then?

Either way - I'll give it a try.


Alex

Re: perf top broken on ppc64

[Ian Munsie]

> > I'm using 32 bit userland and 64 bit kernel on a PowerPC box and it's working for me.
> > Are you building perf from the tip tree?
> 
> I'm using kvm.git which is pretty close to tip. The version says something 2.6.34-rc3'ish. Has anything significantly changed since then?
> 
> Either way - I'll give it a try.


Hey Alex,

Did perf from the tip tree work for you?

Cheers,
-Ian

Re: perf top broken on ppc64

[Alexander Graf]

Am 27.04.2010 um 02:30 schrieb Ian Munsie <imunsie@au1.ibm.com>:

>>> I'm using 32 bit userland and 64 bit kernel on a PowerPC box and  
>>> it's working for me.
>>> Are you building perf from the tip tree?
>>
>> I'm using kvm.git which is pretty close to tip. The version says  
>> something 2.6.34-rc3'ish. Has anything significantly changed since  
>> then?
>>
>> Either way - I'll give it a try.
>
>
> Hey Alex,
>
> Did perf from the tip tree work for you?

Howdy,

Sorry - I left for vacation without replying :).

I tried perf from tip with kvm.git kernel and that did show the same  
effects. I didn't get around to also try to run the tip kernel, and  
won't be within this week :o


Greetings from Gran Canaria,

Alex

>