[Ksummit-2012-discuss] [ATTEND] performance/scalability testing

Alex Shi alex.shi at intel.com
Fri Jun 29 02:54:54 UTC 2012

On 06/28/2012 07:10 PM, Konrad Rzeszutek Wilk wrote:

> On Thu, Jun 28, 2012 at 12:32 AM, Alex Shi <alex.shi at intel.com> wrote:
>> I am working on Intel LKP(Linux kernel performance) project for 5 years.
>> I'd like to attend this meeting to join the discuss of kernel
>> performance/scalability with Andi Kleen and Herbert Xu, I may bring
>> ideas from testing view point.
>> I like to share/discuss the following list if some one have interesting:
>> 1, I like to share what we found or fixed performance issues
>> in upstream kernel recent years.
> That would be neat. Are the fixed issues rather mundane things
> (like using a O(n^2) operations instead of better ones, or blowing away
> caches?) Or are there more deeper than that touching multiple subsystems?

Hard to meet such issue. The most frequent issue are some stupid
mistakes or some careless dependent on other subsystems. but depended
subsystem is not as strange as expected.

>> 2, I like to share/discuss performance/scalability testing technical
>> with detailed examples, like getting profile from
>> systemtap/perf/oprofile/; getting precise time cost for specific kernel
>> path; getting hardware profile from CPU PMU. These info may generated by
>> generic benchmarks: specjbb/oltp/netperf, or by specific
>> micro-benchmarks if generic benchmark is too heavy to introduce other
>> interferes, or by set up stubs in kernel to measure specific kernel path.
>> 3, I like to share/discuss how to implement a full automatic
>> performance/power testing infrastructure to get kernel performance index
>> on many benchmarks and different platforms. and how to find buggy
>> commitment which trigger a performance/power regression on our
>> platforms. also includes how to tuning and stabilise specific benchmarks.
> This is a big project and it sounds quite awesome!

Yes, it really cost me much much time.

 Thought I am bit
> confused - usually
> when on does perf runs you crank up the CPU (so CPU governor is set to
> performance) - but for power testing I would presume you would set it
> to powersave? Both sound mutually exclusive?

We do performance testing with performance gov, and with ondemand gov
for power testing.

>> Regards
>> Alex
>> _______________________________________________
>> Ksummit-2012-discuss mailing list
>> Ksummit-2012-discuss at lists.linux-foundation.org
>> https://lists.linux-foundation.org/mailman/listinfo/ksummit-2012-discuss

More information about the Ksummit-2012-discuss mailing list