From: Benjamin S. Kirk <benkirk@cf...>  20040514 13:14:40

Sure... I have run into the same sort of problem in my discontinuous Galerkin compressible Navier Stokes code. I introduced FEXYZ<> to handle the problem. This can be thought of as exactly analogous to the Monomial finite element, except the shape functions are defined in terms of the XYZ *physical* coordinates on each physical element instead of in (xi,eta,zeta) space on the reference element. The biggest win here is that reinitializing the face values of the shape functions, normals, JxW, etc... is significantly faster. I made substantial improvements to this class just this week, so you'll need the latest CVS to take advantage of it. The other option (which I have wanted to do for a while, but haven't had time) is to provide this information at the *geometric* element level rather than through the finite element. The idea is to provide something like Point p = elem>area(s); Then p.size() is the area (length) of side s, and p.unit() is the face normal vector. Of course you could just use p directly for normal flux calculations. This would be easy to implement since you can do it low in the Elem hierarchy (e.g. for Quad, Tri, Hex, Tet, Pyramid, Prism) and let inheritance take care of the rest. In general, how is libMesh working for your FV stuff? Thanks, Ben Buffat Marc wrote: > Hello Ben, > For my problem running libmesh+my FV code on a linux cluster with RH 7.2 > i will wait for an upgrade to RH 9.0, that will be installed at the end of the > month. > I am doing now some profiling of my FV code with libmesh, PETSC (GMRES+MG) > > Using libmesh profiling I found that 30% of the time was spend in one libmesh > routine: the reinit function of FEBASE class, call on each face of an element > in order to compute the surface, the normal and the centroid of the face. > Can I use another method to get these informations (I am using Monomial > element) and I dont need the shape function and the gradient on the surface > I only need the geomtric information > Regards, > Marc > > 