On Sat, May 22, 2010 at 11:07 AM, jtg <hi.jo...@gmail.com> wrote: > 1) Is there an example in the FiPy 2.1 release that demontrates a > performance improvement from using the trilinos solvers in parallel? > (That is, is there an example that runs in a certain amount of time > with the pysparse solvers and then less time using mpirun and the > trilinos solvers?) I'm hoping to compare the set-up for something > that "works" with what we are doing in order to look for clues as to > the difference.
Most of the examples that use grid meshes should give a speed up. I will run one of the examples and give you some numbers. I'll try and do this within the week. > 2) Just in a very general way, what features would you expect a FiPy > problem to have that would lend themselves to improved performance > under the parallelization scheme FiPy 2.1 implements? Obviously, you have to be using one of the grid meshes and have the longest axis along the x, y or z direction in 1, 2 or 3D. The current partitioning scheme is 1D in all dimensions and utterly trivial and suboptimal, but it worked for the problems that I was interested in (square grid). Our hope is to use PyMetis or Gmsh to implement optimal partitioning. The parallel changes to the code are independent of the partitioning scheme so this shouldn't require major changes just to the Gmsh mesh classes. Hopefully will be part of the next release, but no guarantees right now. -- Daniel Wheeler