Beside theory and experiment, computational science established itself
as a third valuable mode of scientific investigation. We gain more and
more insights from very large and detailed simulations of natural
phenomena. But it is a long way from the physical modeling of these
phenomena, resulting often in differential equations, over the numerical
discretization, the algorithm design, the software implementation, down
to the execution on the hardware level. Many scientist work on the
optimization of the different links of this chain, but local
optimizations without further going considerations do not necessarily
result in an overall improvement. This talk will highlight some of the
misconceptions and point towards more integrative solutions.