On Tue, May 8, 2012 at 6:58 PM, Gideon Simpson <[email protected]> wrote: > While I've used petsc for some larger numerical linear algebra projects > (typically on the order of 10000 degrees of freedom), I'd thought that > smaller problems, with sizes of <= 100 degrees of freedom could be > well handled by serial algorithms. But petsc obviously has a much > more extensive collection of solvers for sparse systems. I was > wondering, has anyone compared performance between the two > libraries on smaller problems?
I'll wager that writing dense matrix operations using the GSL when the GSL is backed by a good BLAS implementation will win for <= 100 DOF problems. But, I've never compared PETSc vs GSL in any meaningful way (and therefore I'm ignoring any PETSc dense capabilities you might use). Like all small work, I suspect your development time will greatly swamp the running time (by a factor of 10,000? 1,000,000?) for small work. I'd lean towards whatever you're most effective at using when developing. - Rhys
