BogoMIPS no longer supported on PowerPC?

I noticed a recent modification to __delay has changed from using empty
loops to reading timebase.  Thus, breaks the calculation of bogomips,
and makes it terribly low.  Is the bogomips no longer being supported on
powerpc arch?

- Leo

On Wed, 2006-08-23 at 18:25 +0800, Li Yang-r58472 wrote:

I noticed a recent modification to __delay has changed from using empty
loops to reading timebase.  Thus, breaks the calculation of bogomips,
and makes it terribly low.  Is the bogomips no longer being supported on
powerpc arch?

What's there to be 'supported'? The only purpose of bogomips is
calibrating the delay loop, and that has changed now by using timebase,
hence bogomips no longer have a purpose. There was even a patch to
remove it from /proc/cpuinfo but I'm not sure whether that got applied
or not.

johannes

Ok, I rephrase my question as is it no longer needed?  Sometimes people
may use this value as a rough estimation of CPU performance.  There are
people even maintained a list of bogoMIPS of most CPUs.  Maybe it is
still needed for some legacy reasons.  I'm not very sure.  :)   At leave
we need to inform people who do look at this magic number and find it
amazingly low.

Best Regards,
Leo

-----Original Message-----
From: Johannes Berg [mailto:johannes@sipsolutions.net]
Sent: Wednesday, August 23, 2006 6:38 PM
To: Li Yang-r58472
Cc: linuxppc-dev@ozlabs.org
Subject: Re: BogoMIPS no longer supported on PowerPC?
=20
On Wed, 2006-08-23 at 18:25 +0800, Li Yang-r58472 wrote:

quoted

I noticed a recent modification to __delay has changed from using

empty

quoted

loops to reading timebase.  Thus, breaks the calculation of

bogomips,

quoted

and makes it terribly low.  Is the bogomips no longer being

supported on

quoted

powerpc arch?

=20
What's there to be 'supported'? The only purpose of bogomips is
calibrating the delay loop, and that has changed now by using

timebase,

hence bogomips no longer have a purpose. There was even a patch to
remove it from /proc/cpuinfo but I'm not sure whether that got applied
or not.
=20
johannes

On Wed, Aug 23, 2006 at 06:55:26PM +0800, Li Yang-r58472 wrote:

Ok, I rephrase my question as is it no longer needed?  Sometimes people
may use this value as a rough estimation of CPU performance.  There are
people even maintained a list of bogoMIPS of most CPUs.  Maybe it is
still needed for some legacy reasons.  I'm not very sure.  :)   At leave
we need to inform people who do look at this magic number and find it
amazingly low.

Bogomips has never had anything to do with performance on PPC, it's just
a number that's based on the decrementer frequency. For processors with
external/fixed decrementer clock, it will not vary between different
processors in a meaningful way.

Every now and then someone will ask "why does my PPC box have such low
bogomips, is it really that much slower than my PC?". Removing all output
would certainly help avoid that confusion.

For reference, my G5 here has 66.56 bogomips, which corresponds well to
the 33MHz decrementer clock it uses.


-Olof

Johannes Berg wrote:

On Wed, 2006-08-23 at 18:25 +0800, Li Yang-r58472 wrote:
 > I noticed a recent modification to __delay has changed from using empty
 > loops to reading timebase.  Thus, breaks the calculation of bogomips,
 > and makes it terribly low.  Is the bogomips no longer being supported on
 > powerpc arch?

What's there to be 'supported'? The only purpose of bogomips is
calibrating the delay loop, and that has changed now by using timebase,
hence bogomips no longer have a purpose. There was even a patch to
remove it from /proc/cpuinfo but I'm not sure whether that got applied
or not.

I missed the patch for this, but I have a related question.
What is the status of the calibration code now for PPC?
Is it just skipped?

I assume that the ability to set a preset loops_per_jiffy
is no longer relevant?
 -- Tim

=============================
Tim Bird
Architecture Group Chair, CE Linux Forum
Senior Staff Engineer, Sony Electronics
=============================

On 8/24/06, Tim Bird [off-list ref] wrote:

Johannes Berg wrote:

quoted

On Wed, 2006-08-23 at 18:25 +0800, Li Yang-r58472 wrote:
 > I noticed a recent modification to __delay has changed from using empty
 > loops to reading timebase.  Thus, breaks the calculation of bogomips,
 > and makes it terribly low.  Is the bogomips no longer being supported on
 > powerpc arch?

What's there to be 'supported'? The only purpose of bogomips is
calibrating the delay loop, and that has changed now by using timebase,
hence bogomips no longer have a purpose. There was even a patch to
remove it from /proc/cpuinfo but I'm not sure whether that got applied
or not.

I missed the patch for this, but I have a related question.
What is the status of the calibration code now for PPC?
Is it just skipped?

The calibration is still used.

I assume that the ability to set a preset loops_per_jiffy
is no longer relevant?

However, loop is no longer true loop now.  It should be called
timebase_change_per_jiffy. :)

- Leo

On 8/23/06, Olof Johansson [off-list ref] wrote:

On Wed, Aug 23, 2006 at 06:55:26PM +0800, Li Yang-r58472 wrote:

quoted

Ok, I rephrase my question as is it no longer needed?  Sometimes people
may use this value as a rough estimation of CPU performance.  There are
people even maintained a list of bogoMIPS of most CPUs.  Maybe it is
still needed for some legacy reasons.  I'm not very sure.  :)   At leave
we need to inform people who do look at this magic number and find it
amazingly low.

Bogomips has never had anything to do with performance on PPC, it's just
a number that's based on the decrementer frequency. For processors with
external/fixed decrementer clock, it will not vary between different
processors in a meaningful way.

Every now and then someone will ask "why does my PPC box have such low
bogomips, is it really that much slower than my PC?". Removing all output
would certainly help avoid that confusion.

For reference, my G5 here has 66.56 bogomips, which corresponds well to
the 33MHz decrementer clock it uses.

I think it's the new way to use timebase or RTC.  Legacy bogomips is
calculated by counting empty loops, which at some degree reflects cpu
performance.  Do you mean it has always been like this for PPC?

- Leo

On Wed, Aug 23, Tim Bird wrote:

I missed the patch for this, but I have a related question.

http://ozlabs.org/pipermail/linuxppc-dev/2006-July/024708.html

On Wed, Aug 23, 2006 at 09:51:38AM -0500, Olof Johansson wrote:

For reference, my G5 here has 66.56 bogomips, 

H3y d00d3, 1ik3, my 12-year-old 486DX100  h4s m0r3 bogomips 
th4n th4t! 486 r0x and p0w3rpc sux!