Help learning from my mistakes

Nicholas Krause xerofoify at gmail.com
Sun Apr 5 15:15:12 EDT 2015



On April 5, 2015 3:05:18 PM EDT, Kristof Provost <kristof at sigsegv.be> wrote:
>On 2015-04-05 14:45:28 (-0400), Nicholas Krause <xerofoify at gmail.com>
>wrote:
>> 
>> 
>> On April 5, 2015 8:29:47 AM EDT, Kristof Provost <kristof at sigsegv.be>
>wrote:
>> >On 2015-04-05 01:15:31 (-0400), Valdis.Kletnieks at vt.edu
>> ><Valdis.Kletnieks at vt.edu> wrote:
>> >> On Sat, 04 Apr 2015 19:00:42 -0400, Nicholas Krause said:
>> >> >  are now slower by up to 3 times,
>> >> 
>> >> So what debugging did you already do to try to narrow this down?
>> >>
>> >While we're on the subject, let me indulge a pet peeve: How did you
>> >measure this? Can you reproduce the measurement? Is the difference
>> >statistically significant? What aspect of performance is worse? Task
>> >switching time? IO throughput?
>> >
>> >Benchmarking is *hard*, and 'up to 3 times' is not the result of a
>> >careful, well thought out benchmark.
>> >
>> I ran the make command on all of my CPU cores and timed  it. 
>Furthermore it seems to be slower up to 3 times then my distribution
>kernel based off 3.16.
>
>So, in other words, no, you've not done a careful benchmark at all.
>
>What's the standard deviation on both measurements? How confident are
>you (your answer should be a number) that there is a difference?
>How have you accounted for VFS cache effects? Have you ensured that
>both
>builds were from a completely clean tree? How have you ensured that no
>other jobs interfered with the build (i.e. cron, user interaction,
>...)?
>
>Once you've established that there really is a difference you need to
>start isolating contributing factors and figure out what's causing the
>difference. Alternatively, you can start trying to figure out which
>part
>of overall system performance has been degraded. That requires actually
>understanding how the system works and using inspection tools (top,
>iostat, perf, ...) to find out where the performance bottleneck is.
>
>None of that is easy or quick. None of these questions can be answered
>in one quick sentence. Certainly your overall answer should include
>more
>information, and demonstrate more thought and effort, than 'I did a
>build.'.
>
>Regards,
>Kristof
 The only difference I can tell is the kernel version.  I was wondering however if anyone who has more experience in the kernel can give me tips on how to address my benchmarking. 
Thanks, 
Nick

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.



More information about the Kernelnewbies mailing list