Mailing List Archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [tlug] CPU speed: reported vs. actual

Darren Cook writes:

 > To answer you and Attila, my main data analysis/search work is
 > CPU-bound;

I'll take your word for that, but I don't think GFLOPS is an
appropriate way to benchmark search, and it may not be all that
relevant to data analysis (for example, one of my statistician
colleagues spends a lot of time in random()).

 > if I'm missing a simple configuration step to have it complete in
 > half the time it is worth an investment of my time to discover it.

Sure.  That's not what I'm questioning though; I'm questioning the
benchmarks you're using to measure (a) the theoretical maximum and
(b) the actual performance in your application.  My experience is that
the benchmarks that vendors use tend to be factor-of-2 optimistic, and
the 40% you're seeing is close to that.

 > I tried installing and using cpufreq-set as Mattia suggested, with no
 > improvement; next thing to try is booting into Windows and running
 > Intel's linpack and some other benchmarks from there.

Do you mean a binary of LINPACK?  If so, there's no guarantee that you
can get anywhere close to the performance of Intel's carefully-tuned-
for-Windows-and-the-CPU build in real applications.  You really need
to look at the configuration, and see if its settings can be applied
to your own work.

Home | Main Index | Thread Index

Home Page Mailing List Linux and Japan TLUG Members Links