Supercomputers allow the simulation of very large models both with
respect to realism (number and complexity of physical processes
involved) and accuracy (number of degrees of freedom). Due to the
complexity of the models the life-cycle of numerical software is
typically much longer than that of supercomputer hardware. This
requires that numerical software is flexible and extensible on the one
hand but also efficient with respect to run-time and memory on the
other hand. The software framework DUNE (Distributed and Unified
Numerics Environment), developed by groups in Freiburg, Berlin and
Heidelberg, tries to achieve this through a strict separation of data
structures and algorithms, use of generic programming techniques and
reuse of existing software. In the first part of the talk I will
present relevant aspects of this software and some numerical results
while the second part of the talk concentrates on scalability
aspects of a structured mesh multigrid code based on the DUNE
framework.