Projekt

Supercomputer architecture is moving quickly to multi-core and many- core architectures. An additional trend is the increasing use of special purpose accelerators, e.g. in form of graphics cards, the Cell processor, or reconfigu- rable hardware. This has the potential to deliver unprecedented performance at lower cost and reduced power consumption. However, this trend opens many unanswered questions on how these devices can be use effectively in real life supercomputing applications, since these accelerators are only effec- tive on certain algorithmic structures that exhibit a high degree of regularity and they are inefficient or very difficult to program for others. At the LSS we have developed (partly with KONWIHR support) application software, whose basic structure supports heterogeneous architectures. For this project we will focus on the waLBerla framework for simulating complex flows that have a wide range of applications [4, 5]. The software waLBerla is designed such that it distinguishes between modules that require complex algorithmic logic and regular structures that can exploit a very high degree of parallelism. While the latter parts are very well suited for a potential use of multi-core parallelism and accelerator technologies, the algorithmically more complex parts also can be run on the conventional parts of a heterogeneous system. We therefore propose to extend waLBerla in order to exploit current and future heterogeneous multi-core architectures.