On Apr 1, 2009, at 3:28 AM, Sum Thai Wong wrote:
I was trying to run some rough benchmarks to compare the speedup using Trilinos. To my surprise, Fipy actually runs faster without Trilinos, and by a significant margin. Could this be possible?
Possible and probable. If you look at the results for "sandbox" at http://matforge.org/fipy/build/trunk you can see that we get the same thing (sandbox is the only test slave that's properly configured for Trilinos at the moment).
The test suite is not actually of much value for benchmarking. It's running through a lot of very small problems, so the overhead in both memory and speed is substantial. It doesn't surprise me that the overhead for using Trilinos is even worse.
I am presently working on setting up a benchmarking script for our own ongoing diagnostics. I'll post when I have the results up.
I notice that only a single core is being utilised?
Correct. Presently, the only advantage of using Trilinos with FiPy is that has a much richer suite of preconditioners and, under some circumstances, can solve the matrix in parallel to some advantage. Because FiPy can only presently *build* the matrix in serial, this really comes down to a case-by-case situation. Dan Wheeler is more experienced with this, but I think he's out for a few days.
We are working on a couple of routes to parallelizing the whole process and will post here when we have something available.
Thanks for your interest and let us know how we can further help.
