Benchmarking scientific packages for general and architecture specific builds
Description of the project: For Debian packages we need to strip optimisation options from upstream build system like for instance -march=native to make sure that the code will run also on those computers who are different from the one the package was actually built. There are reports by Debian Science members that the effect of those optimisation options is quite low and extra compilation on local machines do not result in a sensible gain of speed. The task for the student will be to do sensible performance tests of packages to get some valuable data to decide whether it makes sense to create new means to support users in doing local builds with better optimisation. Bonus points for the student if it turns out that optimised local builds are sensible to implement an easy way for the user to do so.
Confirmed Mentor: Name of the mentor
How to contact the mentor: (mail, IRC, etc)
Confirmed co-mentors: Andreas Tille <tille@debian.org>, Lumin <cdluminate@gmail.com>
Deliverables of the project: Comparison of real life examples with general Debian package and optimised local build
Desirable skills: Scientific background, comfortable with linear algebra (atlas, blas), comfortable with gcc and optimisation
What the intern will learn: The student will get an introduction into the Debian Science team and the team maintained packages. Some basic packaging skills will be teached if needed. The project is also supported by members of Debian Astro, Debian Med, ?DebiChem teams basically everywhere where high performance computing in science is needed.
Application tasks: Contact Debian Science list <debian-science@lists.debian.org> to discuss appropriate example cases to test. Please do all conversation via this list and add the tag [GSoC] to mail subject.
Related projects: DebianScience, Debian Astro, Debian Med