Collection of abstracts 14th
GAMM-Seminar Kiel on
|
Parallelization of state-of-the-art numerical software (e.g., multigrid methods on unstructured grids with adaptive refinement) is a tedious task which often requires great efforts for development and maintenance of the resulting codes. Through the use of high-level parallel programming models (e.g. DDD, the Dynamic Distributed Data library, University of Stuttgart) parallelisation work can be lifted onto a more abstract level, which leads to hierarchical software packages. However, lots of implementation details and therefore sources of possible errors remain.
In this talk, a powerful concept of automatic parallel source code generation based on formal specifications is presented. By using techniques from research in Artificial Intelligence like automatic theorem proving, planning or generate & test schemes efficient algorithms can be constructed, which map the data structures of the sequential numerical software onto distributed graphs. Therefore, most of the parallelisation work is replaced by choosing global properties of the parallel program and its data in a specification database. Moreover, the concept is also applicable to a broad range of problems different from parallelisation including mesh refinement/de-refinement, code verification, automatic debugging or hardware-specific problems like vectorisation or cache-optimization.
Mail to WebMaster |
[Wed Dec 17 10:55:09 MET 1997] |
Impressum |