Domain decomposition methods for large linearly elliptic three-dimensional problems (Q2640331)

The authors are interested in solving of large three-dimensional problems drawn from linear elasticity theory. The method of solution is to divide the domain \(\Omega\) of the problem into several subdomains \(\Omega_ i\). The subdomains abut along internal boundaries, the collection of which is called \(S\). The overall iteration can be described in the following steps: 1. Given values \(\lambda\) on \(S\), solve Dirichlet problems to yield functions \(u_ i\) on each \(\Omega_ i\). It is not expected that \(\partial u_ i/\partial n+\partial u_ j/\partial n=0\) along the common internal boundary of \(\Omega_ i\) and \(\Omega_ j\). 2. Solve Neumann problems for functions \(\phi_ i\) on \(\Omega_ i\) using as boundary data a linear combination of normal derivatives of solutions from Step 1. 3. Update \(\lambda\) using the functions \(\phi_ i\). The authors prove that the iteration converges at least linearly, with a constant independent of the mesh size on individual subdomains. In their discussion, the authors refer to the Neumann step above as a ``preconditioning'' step, but it is hard to see how to proceed without it. The authors discuss application of the method to two and three dimensional problems for the Laplace operator and for a problem arising from elastic analysis of a TRIFLEX connection arm made of a glass resin elastomere composite material. They discuss implementation issues associated with the architecture of the CRAY 2 and Intel Hypercube computers. Different decompositions as well as data structures are considered, and timing comparisons are made.