Re: [PATCH 7/7] Simple Performance Counters: SLUB instrumentation

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Christoph Lameter (clameter@sgi.com) wrote:
> On Fri, 17 Aug 2007, Mathieu Desnoyers wrote:
> 
> > Why do you printk inside the timing period ? Filling the printk buffers
> > or outputting on things such as serial console could really hurt your
> > results.
> 
> It was easier to code?
> 
> > I hope you run your system with idle=poll and without frequency scaling
> > at all, because otherwise your cycle count would be completely off on
> > many AMD and Intel CPUs. You can have a look at this (very rough)
> > document on the topic: 
> 
> The cpu will definitely not be idle during these measurements and no 
> frequency scaling is active.

The problem is that if the cpu is idle _before_ the measurements, the
frequency will change differently from one cpu to another. Therefore,
the cycle counters may have large offsets when you start your tests. So,
if get_cycles() is executed on different CPUs (thread being migrated)
between the beginning and the end of the test, the results would be
skewed.

Mathieu


-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68