Re: Building octave 3.8.2 from source on Ubuntu 14.04 for using multiple cores

[Sunil Shah]

Hi ST,

Thanks for pointing out the obvious error on my part in testing openblas.  

Happily, with octave 3.8.1 switching to 

/usr/lib/openblas-base/libblas.so.3     

and 

/usr/lib/lapack/liblapack.so.3

does work.   It also uses all available cores for EC2.  So, now I see cpu usage go up to 1600% on 16 cores. 

When building octave 3.8.2 from source, do I need to specify these blas and lapack libraries as configure options explicitly or it will pick them automatically given that they have been set through update-alternatives ?

If I do need to specify libraries through configure options, is it better to specify the dynamic libraries or static? I am using no other configure options. 

Sunil

On Fri, Sep 5, 2014 at 2:23 AM, <address@hidden> wrote:

On Friday, September 05, 2014 03:02 AM, Sunil Shah wrote:

Hi ST,

Sorry for the incomplete post.

Please use bottom post. It's easier to follow.

Following up on the link on R you sent:

Using octave 3.8.1 binary package on 14.04 I can switch to atlas and lapack.

/usr/lib/atlas-base/atlas/libblas.so.3

/usr/lib/atlas-base/atlas/liblapack.so.3


However, it does not enable octave to use multiple cores for standard
dense matrix test on lu(randn(5e3))

It is faster than stock blas and atlas.


Curiously, unlike the link on R, I can not use
/usr/lib/openblas-base/libblas.so.3 with
/usr/lib/atlas-base/atlas/liblapack.so.3


I get this error message

/usr/bin/octave-cli: symbol lookup error: /usr/lib/liblapack.so.3:
undefined symbol: ATL_idamax

You seemed to be mixing the lapack provided by atlas with the blas provided by openblas.

Please see the Debian wiki page that I mentioned earlier. It has a section on selecting the appropriate lapack package. The link is repeated here:

https://wiki.debian.org/DebianScience/LinearAlgebraLibraries

How to run octave faster in the cloud would certainly be interesting to people like me who has no such experience.

If your are running octave locally on an x86 machine, I'd suggest you remove all atlas related stuff and install openblas. Just read the README.Debian that comes with it as suggested in the Debian wiki and follow the instructions of re-building openblas locally. Most likely your cpu will be detected correctly.

Regards,
ST
--