time accounting problem (powerpc only?)

time accounting problem (powerpc only?)

[Johannes Berg]

Hi,

On my powerbook, with NO_HZ and HIGH_RES_TIMERS, I observed recently
that powernowd would not ever switch between CPU speeds.

Doing some scripting to read /proc/stat every half second and print the
differences, I get output like this on a mostly idle system:

# for reference:
# [user, nice, system, idle, iowait, irq, softirq, steal, guest]

[4, 0, 3, 46, 0, 264, 0, 0, 0]
[5, 0, 4, 41, 0, 266, 0, 0, 0]
[4, 0, 7, 37, 0, 299, 0, 0, 0]

Now starting an empty while (1) loop, I see:

[53, 0, 1, 0, 0, 243, 0, 0, 0]
[53, 0, 2, 0, 0, 224, 0, 0, 0]
[45, 0, 7, 0, 0, 249, 0, 0, 0]

As you can see, the "irq" time is very high so that "user" (mostly my
CPU burning program) only accounts for maybe 15%. This leads to
powernowd thinking that all is fine and no switching is required. Also,
the whole stuff leads to top(1) displaying about 80-90% "hi" (hard irq?
calculated as irq - softirq?) time.

Does anybody have an idea why the irq time is so high? And no, I don't
think it can actually be true, I'm seeing maybe 500 interrupts/second
(most due to USB).

johannes

Re: time accounting problem (powerpc only?)

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 207 bytes --]

Contrary to what I claimed later in the thread, my 64-bit powerpc box
(quad-core G5) doesn't suffer from this problem.

Does anybody have any idea? I don't even know how to debug it further.

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 828 bytes --]

Re: time accounting problem (powerpc only?)

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 200 bytes --]


> On my powerbook, with NO_HZ and HIGH_RES_TIMERS, I observed recently
> that powernowd would not ever switch between CPU speeds.

Also happens without NO_HZ (with HIGH_RES_TIMERS).

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 828 bytes --]

Re: time accounting problem (powerpc only?)

[Tony Breeds]

On Mon, Nov 26, 2007 at 05:23:13PM +0100, Johannes Berg wrote:
> Contrary to what I claimed later in the thread, my 64-bit powerpc box
> (quad-core G5) doesn't suffer from this problem.
> 
> Does anybody have any idea? I don't even know how to debug it further.

I'll see if I can grab an appropriate machine tomorrow and have a look at
it.  I think it's just an accounting bug, which is probably my fault :)

Yours Tony

  linux.conf.au        http://linux.conf.au/ || http://lca2008.linux.org.au/
  Jan 28 - Feb 02 2008 The Australian Linux Technical Conference!

Re: time accounting problem (powerpc only?)

[Paul Mackerras]

Johannes Berg writes:

> Contrary to what I claimed later in the thread, my 64-bit powerpc box
> (quad-core G5) doesn't suffer from this problem.
> 
> Does anybody have any idea? I don't even know how to debug it further.

I see it on my G4 powerbook.  However, a compute-bound task runs just
as fast (i.e. completes in the same elapsed time) as on older kernels,
so it does look like it is just an accounting problem.  If the CPU was
really spending more than half its time servicing hardware interrupts,
the task should take more than twice as long to complete.

Paul.

Re: time accounting problem (powerpc only?)

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 727 bytes --]


On Tue, 2007-11-27 at 20:57 +1100, Paul Mackerras wrote:
> Johannes Berg writes:
> 
> > Contrary to what I claimed later in the thread, my 64-bit powerpc box
> > (quad-core G5) doesn't suffer from this problem.
> > 
> > Does anybody have any idea? I don't even know how to debug it further.
> 
> I see it on my G4 powerbook.  However, a compute-bound task runs just
> as fast (i.e. completes in the same elapsed time) as on older kernels,
> so it does look like it is just an accounting problem.  If the CPU was
> really spending more than half its time servicing hardware interrupts,
> the task should take more than twice as long to complete.

Exactly. So where do I look for the accounting bug?

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 828 bytes --]

Re: time accounting problem (powerpc only?)

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 547 bytes --]


On Tue, 2007-11-27 at 16:00 +1100, Tony Breeds wrote:
> On Mon, Nov 26, 2007 at 05:23:13PM +0100, Johannes Berg wrote:
> > Contrary to what I claimed later in the thread, my 64-bit powerpc box
> > (quad-core G5) doesn't suffer from this problem.
> > 
> > Does anybody have any idea? I don't even know how to debug it further.
> 
> I'll see if I can grab an appropriate machine tomorrow and have a look at
> it.  I think it's just an accounting bug, which is probably my fault :)

Thanks. Let me know if you need any info.

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 828 bytes --]