Peak performance of computer systems is exponentially growing since decades. In accordance with Moore’s law the Pflop/s level will be reached within a few years. However, the performance gap between peak performance and sustained performance is continually growing. Ever increasing effort is needed to utilize top performance computers satisfactorily. 

A possible answer to meet this challenge is a multilayered approach that starts on the highest level of abstraction, i.e., mathematical models and numerical methods tailored to the needs of modern computer architecture, continues with automatic code generation, adaptation, and optimization, and finally goes down to special purpose compilers, which utilize problem specific information underlying the code to be translated. 

Examples will be given in the context of IBM’s Blue Gene systems, reaching up to a 60 Tflop/s code.