From: Wolfgang Bangerth <bangerth@ma...>  20080124 15:57:44

> I am currently looking into designing a new 3D finite element > geoelectric code. I have come across deal.II and libmesh. Both > libraries look quite promising, each having their > advantages/disadvantages. On the plus side deal.II supports Nedelec > elements which I am quite interested in using. However, libmesh allows > to distribute the degrees of freedom (in addition to parallelising the > linear algebra). A major design factor is memory requirements. This also > seems to be one of the main distinguishing factors of the two libraries. > Do you have estimates for the memory requirements for the stiffness > matrix and load vector? That depends on the problem you solve. For the Laplace equation in 2d with Q1 elements you have 9 entries per row of the matrix, so you are talking about 108 bytes in the matrix per degree of freedom. Plus storage for solution vector, right hand side, and temporary vectors in the CG method. For the Laplace equation in 2d, the linear system is not the largest memory consumer of a finite element program, deal.II needs several hundred (but less than 1000) bytes per degree of freedom. That changes if you go to 3d and systems of equations because there you have many more entries per row of the matrix (for example for an equation like 'curl curl B = f' you would have 81 entries per row of the matrix already for the lowest order elements. So you would already consume ~1000 bytes per degree of freedom for the linear system. This is independent of the library you use, unless you go to schemes where the matrix is not stored. Higher order elements would use even more memory, of course. In these cases, the memory footprint of the linear system is larger than grid, DoF handler, and the rest together. > What size problems can be reasonably run? On a machine with 16GB, using deal.II, you can solve linear systems with a few million degrees of freedom (in 2d) or about a million (in 3d) without too much trouble. On clusters we have solved problems with >20 million unknowns. Now let's see what our friends from the competition have to say ;) Cheers W.  Wolfgang Bangerth email: bangerth@... www: http://www.math.tamu.edu/~bangerth/ 