Re: [PATCH] powerpc/mm: Add trace point for tracking hash pte fault

["Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>]

Michael Ellerman <mpe@ellerman.id.au> writes:

> On Wed, 2015-01-21 at 14:15 +0530, Aneesh Kumar K.V wrote:
>> Michael Ellerman <mpe@ellerman.id.au> writes:
>> 
>> > On Tue, 2015-01-20 at 17:05 +0530, Aneesh Kumar K.V wrote:
>> >> This enables us to understand how many hash fault we are taking
>> >> when running benchmarks.
>> >> 
>> >> For ex:
>> >> -bash-4.2# ./perf stat -e  powerpc:hash_fault -e page-faults /tmp/ebizzy.ppc64 -S 30  -P -n 1000
>> >> ...
>> >> 
>> >>  Performance counter stats for '/tmp/ebizzy.ppc64 -S 30 -P -n 1000':
>> >> 
>> >>        1,10,04,075      powerpc:hash_fault
>> >>        1,10,03,429      page-faults
>> >> 
>> >>       30.865978991 seconds time elapsed
>> >
>> > Looks good.
>> >
>> > Can you attach some test results that show it's not hurting performance when
>> > it's disabled.
>> 
>> ebizzy with -S 30 -t 1 -P gave
>> 13627 records/s -> Without patch
>> 13546 records/s -> With patch with tracepoint disabled
>
> OK. So that's about -0.6%. Are we happy with that? I'm not sure.
>
> Can you do a few more runs and see if that's a stable result.

That is within the run variance for that test. Infact I found it
difficult to get a stable records/s with ebizzy run, even after fixing
the random state variable and forcing single thread. I ended up doing
a micro benchmark that allocate a large region and touch one byte per
page.

That resulted in
time perf stat -e page-faults -e powerpc:hash_fault  ./a.out

 Performance counter stats for './a.out':

         10,00,062      page-faults                                                 
         10,00,068      powerpc:hash_fault                                          

      12.414350121 seconds time elapsed


real    0m12.558s
user    0m0.577s
sys     0m11.932s

and with that test we have an average system time for 10 run
Without patch
sys: 0m11.2425

With patch:
sys: 0m11.3258

ie, a -0.7% impact 

If that impact is high we could possibly put that tracepoint within #ifdef
CONFIG_DEBUG_VM ?

-aneesh