This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] Porting string performance tests into benchtests


On 04/05/2013 01:38 AM, Siddhesh Poyarekar wrote:
> On Wed, Apr 03, 2013 at 11:40:42PM -0400, David Miller wrote:
>> I really want to see on the cpu cycle level whether the changes I make
>> to the pre-loop and post-loop code make any difference.
>>
>> And on sparc chips I don't have the issues that can make the cpu cycle
>> counter inaccurate or less usable as a timing mechanism.
> I realized that my understanding of CLOCK_PROCESS_CPUTIME_ID was
> flawed.  While it is described as 'cpu time consumed by the process',
> it seems to still be sufficiently impacted by system load.  Maybe the
> cost of switching gets added as well, I'm not sure.  What I'll do now
> is see if HP_TIMING gives reasonably consistent results in the same
> conditions as CLOCK_PROCESS_CPUTIME_ID does and if it does, I'll
> modify the benchmark code to use it if available.  If not, we fall
> back to CLOCK_PROCESS_CPUTIME_ID.
>
> Does that sound like a good plan?
>
> Siddhesh
>
For PowerPC I would prefer to use HP_TIMING, since it uses the timebase mechanism
and it have a higher resolution, does not require a a syscall or even a vDSO
access, and it is not susceptible to CPU scaling.

I'm not sure if it is worth, but maybe it would be useful to add a configurable
timestamp calculation. And related either timestamp way will be impacted by the
system lead at the time: os jitter should be controlled or minimized by the
tester to get good measurements.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]