[Starcluster] NumPy and SciPy tests fail on Ubuntu 9.10 Karmic

[Andrew Straw]

Justin Riley wrote:
> I'll give stdeb a shot and hopefully create some custom numpy/scipy debs for 
> the ami. However, the harder package to customize I suspect will be lapack 
> which is pretty much the entire reason for custom compiling numpy/scipy debs 
> in the first place. I noticed you have or had a repo for ubuntu numpy/scipy 
> (from http://www.scipy.org/Installing_SciPy/Linux). Did your packages use the 
> default debian lapack?
>   
I have always used the default lapack/atlas -- linear algebra has never
been a performance bottleneck for me. So, IIRC, my repositories have
never had custom lapacks.

My feeling is that even if you don't manage to .deb-ify your lapack
stuff (do you know about checkinstall?), it is still good to .deb-ify
numpy/scipy just so that "apt-get install python-numpy" will do the
right thing (i.e. not install the Ubuntu repo version if it has a lower
version number than your custom install). This is somewhat independent
from whether "apt-get install liblapack" does clobber your custom lapack
installation. Ideally, though, your custom numpy would have a "Depends:
custom-lapack ( >= custom-version-number )" in the debian/controls file
so that these two packages' dependency graph would be known by the apt
system. This, would, of course, require getting a .deb for the custom
lapack, as well.

Anyhow, I am happy to attempt to what knowledge I have of debian/ubuntu
packaging arcana -- I am moving a lot of my analysis stuff to SGE and
the plan is that some of this will run on EC2/starcluster.

-Andrew