Why this Great 'SpeedUp'?

6 views (last 30 days)
andre barbieri
andre barbieri on 11 Oct 2017
Commented: andre barbieri on 12 Oct 2017
Error or SuperHiperPower optimization???
  6 Comments
David Goodmanson
David Goodmanson on 11 Oct 2017
thanks Cedric
OCDER
OCDER on 11 Oct 2017
Hi Andre, I'm just a volunteer trying to answer questions or point towards resources to help people get answers.
Your question "Error or SuperHiperPower optimization???", would yield an Error vs Optimization choice answer, which probably won't satisfy your curiousity. The answer would be optimization since no errors are being returned.
If you want to know "What did MathWorks do to make lu faster in 2017b than in 2017a?", then we can try a bit more to help. BUT, turns out lu is a built-in compiled function with protection against viewing the source code optimization... Wish we could help you more by investigating changes made, but seems like we can't:

Sign in to comment.

Answers (1)

Steven Lord
Steven Lord on 11 Oct 2017
First note that you should not compare the results of running different versions of bench on different releases as bench may very well change from release to release. We mention that in the help for bench as well as at the bottom of the figure containing the numeric results of the bench run. [I don't know offhand if it did for release R2017b or not.] If you must compare two releases, run the older version of bench on both releases.
Even if the LU task did not change between release R2017a and R2017b the machine on which it was run appears to have changed. In R2017a the machine claiming the top spot in the table was listed as a Windows 7 machine; in R2017b it was listed as a Windows 10 machine.
  3 Comments
Jan
Jan on 12 Oct 2017
@andre: The code of bench is subject to changes. You have to run the same bench.m file in both Matlab versions for a comparison. The 2016b version of the LU test is:
n = 2400;
A = randn(n, n);
tic
B = lu(A);
t = toc;
What do you see in the bench functions for 2017a and b? If they are equal, this would mean that LU uses faster methods: More cores or AVX code?
andre barbieri
andre barbieri on 12 Oct 2017
'1.46x speedup' for LU only in the First Cpu (same cpu, just change 2017a to 2017b) !!! Think that this was an error..

Sign in to comment.

Products

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!