Re: Performance regression testing?

[Mike Stump <mrs at apple dot com>]

On Nov 29, 2005, at 11:51 AM, Laurent GUERBY wrote:
> On Mon, 2005-11-28 at 17:55 -0800, Mike Stump wrote:
>
> > My feeling is that we should have such a suite.  I'd favor a micro
> > style, where we are measuring clock cycles (on machines that can
> > expose them x86/v9), [...]
>
> A while ago I looked at rdtsc on x86-linux, but I couldn't find a way,
> other than statistical / multiple runs and take minimum, to known
> how many context switch happened between two points in the measured
> code. Is this number available in some ways?

Nope, not portably.  An OS is free to save out the tick count for  
user land ticks, and make that available to the program.  The idea  
would be you take the tick counter add in this bias, and then presto,  
you have user ticks for the process/thread.  Just file a bug report  
against your favorite OS for the feature, or ask an OS expert if it  
exists.

When not available, just run 5, or 10 times, and take the minimum and  
don't worry.  interrupts are rare events, they don't happen but 1 in  
a million clock cycles.  :-)