Re: [patch 9/9] Scheduler profiling - Use conditional calls

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

* Andrew Morton (akpm@linux-foundation.org) wrote:
> On Fri, 1 Jun 2007 11:54:13 -0400 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> wrote:
> 
> > * Andrew Morton (akpm@linux-foundation.org) wrote:
> > > On Wed, 30 May 2007 10:00:34 -0400
> > > Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> wrote:
> > > 
> > > > @@ -2990,7 +2991,8 @@
> > > >  			print_irqtrace_events(prev);
> > > >  		dump_stack();
> > > >  	}
> > > > -	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
> > > > +	cond_call(profile_on,
> > > > +		profile_hit(SCHED_PROFILING, __builtin_return_address(0)));
> > > >  
> > > 
> > > That's looking pretty neat.  Do you have any before-and-after performance
> > > figures for i386 and for a non-optimised architecture?
> > 
> > Sure, here is the result of a small test comparing:
> > 1 - Branch depending on a cache miss (has to fetch in memory, caused by a 128
> >     bytes stride)). This is the test that is likely to look like what
> >     side-effect the original profile_hit code was causing, under the
> >     assumption that the kernel is already using L1 and L2 caches at
> >     their full capacity and that a supplementary data load would cause
> >     cache trashing.
> > 2 - Branch depending on L1 cache hit. Just for comparison.
> > 3 - Branch depending on a load immediate in the instruction stream.
> > 
> > It has been compiled with gcc -O2. Tests done on a 3GHz P4.
> > 
> > In the first test series, the branch is not taken:
> > 
> > number of tests : 1000
> > number of branches per test : 81920
> > memory hit cycles per iteration (mean) : 48.252
> > L1 cache hit cycles per iteration (mean) : 16.1693
> > instruction stream based test, cycles per iteration (mean) : 16.0432
> > 
> > 
> > In the second test series, the branch is taken and an integer is
> > incremented within the block:
> > 
> > number of tests : 1000
> > number of branches per test : 81920
> > memory hit cycles per iteration (mean) : 48.2691
> > L1 cache hit cycles per iteration (mean) : 16.396
> > instruction stream based test, cycles per iteration (mean) : 16.0441
> > 
> > Therefore, the memory fetch based test seems to be 200% slower than the
> > load immediate based test.
> 
> Confused.  From what did you calculate that 200%?
> 
> > (I am adding these results to the documentation)
> 
> Good, thanks.


(48.2691-16.0441)/16.0441 = 2.00

Which means that it is 200% slower to run this test while fetching the
branch condition from main memory rather than using the load immediate.

We could also put it like this : the speedup of the load immediate over
the memory fetch is 3.

48.2691/16.0441 = 3.00

Is there a preferred way to present these results in the documentation ?

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