help-octave
[Top][All Lists]
Advanced

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

performances issues


From: Luca Tagliacozzo
Subject: performances issues
Date: Tue, 3 May 2011 21:53:23 +0200

I finally manged to compile octave (both 3.2.4 and 3.4.0) using the
openblas library.
The openblas library is compiled so that it uses up to 8 cores.
For the problem I am solving several svd of matrices of double of size
of 1000x 1000 it uses around 6 cores.
The timing results is a bit unexpected
Using the newly compiled version I get


Nspins =  6
hz =  1
n =  1
ans =  112.94
n =  2
ans =  90.390
cpu_time=203.34
Elapsed time is 40.9 seconds.


While using the precompiled version (which is linked against a
serialized version of atlas precompiled in my distro) I get


Nspins =  6
hz =  1
n =  1
ans =  9.7800
n =  2
ans =  8.0100
cpu_time=17.79
Elapsed time is 18.99 seconds.

While I can understand that the output of cputime measures more time
since it uses more cores in the first instance, it looks very strange
that the elapsed time (measured via tic toc) is 3 times bigger,
does it mean that atlas serialized is faster than gotoblas?

An interesting thing I noticed is that also the version I recently
compiled still links against the blas and lapack atlas libraries even
if it also links to the openblas libraries....
Did I something wrong?


Clearly there is no point in using 5 cores and running 3 times slower
than using one core, I am a bit confused
Does someone has a good explaination of this?

Regards
Luca


reply via email to

[Prev in Thread] Current Thread [Next in Thread]