I am using GSL from another library of my own to perform numerical
integration of vectorial differential equations. After optimizing and
parallelizing most of my library, I ended up with the conclusion that GSL is
a major bottle neck in my computation, simply because it is not parallelized
to exploit multi-core achitectures.
Have you tried solving many such problems in parallel using
threadprivate GSL workspaces? This permits you to put a parallel
section outside of many GSL invocations by ensuring GSL's working
storage is kept local to each thread.
- Rhys