Re: [PATCH 3/3] perf diff: Introduce tool to show performance difference

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Dec 14, 2009 at 09:30:26PM -0200, Arnaldo Carvalho de Melo wrote:
> Em Mon, Dec 14, 2009 at 02:47:08PM -0800, Paul E. McKenney escreveu:
> > On Mon, Dec 14, 2009 at 08:09:31PM -0200, Arnaldo Carvalho de Melo wrote:
> > > From: Arnaldo Carvalho de Melo <acme@redhat.com>
> > > 
> > > I guess it is enough to show some examples:
> > 
> > Very cool!!!
> > 
> > Some questions on the numbers below...
> 
> Lets go!

[ . . . ]

Thank you for the info!!!

> in another, and look at what happen when you flip those xterms.
> 
> And if you want to see an html rendering of what I wanted to get coming
> accross:
> 
> http://esporte.uol.com.br/futebol/campeonatos/brasileiro/2009/serie-a/classificacao.jhtm

Differential profiling applied to soccer statistics...  No -there- is a
scary thought!!!  ;-)

							Thanx, Paul

> > > [root@doppio linux-2.6-tip]# perf diff -p | head -5
> > >    1        +1.00%     /lib64/libc-2.10.1.so   _IO_vfprintf_internal
> > >    2                       [kernel.kallsyms]   __kmalloc
> > >    3    +1             /lib64/libc-2.10.1.so   __GI_memmove
> > >    4    +4             /lib64/libc-2.10.1.so   _int_malloc
> > >    5    +7  -1.00%         [kernel.kallsyms]   __d_lookup
> > 
> > The third column is percent of total execution time?  Or percent change
> > in profile ticks?  My guess is the former.
> 
> counter percentage wrt the total number of hits for that particualr
> session. The unit is whatever is specified in --event, i.e. the counter
> specified, whichever it is.
> 
> > > [root@doppio linux-2.6-tip]# perf diff -v | head -5
> > >    1        361449551 326454971 -34994580     /lib64/libc-2.10.1.so   _IO_vfprintf_internal
> > >    2        151009241 135701435 -15307806         [kernel.kallsyms]   __kmalloc
> > >    3    +1  101805328 105471269  +3665941     /lib64/libc-2.10.1.so   __GI_memmove
> > >    4    +4   78041440 101550435 +23508995     /lib64/libc-2.10.1.so   _int_malloc
> > >    5    +7   59536172  98074985 +38538813         [kernel.kallsyms]   __d_lookup
> > > [root@doppio linux-2.6-tip]# perf diff -vp | head -5
> > >    1        9.00% 8.00% +1.00%     /lib64/libc-2.10.1.so   _IO_vfprintf_internal
> > >    2        3.00% 3.00%                [kernel.kallsyms]   __kmalloc
> > >    3    +1  2.00% 2.00%            /lib64/libc-2.10.1.so   __GI_memmove
> > >    4    +4  2.00% 2.00%            /lib64/libc-2.10.1.so   _int_malloc
> > >    5    +7  1.00% 2.00% -1.00%         [kernel.kallsyms]   __d_lookup
> > 
> > If these examples are all using the same numbers, then the percentages
> > must be of total execution time rather than percent change in the
> > profiling ticks?
> 
> Its all using the same perf.data.old + perf.data files, so the numbers
> are for the default -e metrics, which is PERF_COUNT_HW_CPU_CYCLES.
> 
> > > [root@doppio linux-2.6-tip]#
> > > 
> > > This should be enough for diffs where the system is non volatile, i.e. when one
> > > doesn't updates binaries.
> > > 
> > > For volatile environments, stay tuned for the next perf tool feature: a buildid
> > > cache populated by 'perf record', managed by 'perf buildid-cache' a-la ccache,
> > > and used by all the report tools.
> > 
> > For scalability studies, it would be very cool to have a ratio as well
> > as a difference, but again, good stuff!!!
> 
> Point taken!
> 
> Please let me know about any other issue or suggestion you may come to!
> 
> - Arnaldo