PP-B: Model-based Parallelization
Sequential programs need adaptation for parallelization on high performance computing (HPC) architectures. Parallelization requires expert knowledge across software design, algorithm, code and hardware. Hence, parallelization of legacy software is currently a highly manual task with a lack of hard correctness or performance guarantees.
The goal of this part project is to develop a novel, model-based parallelization approach with extensive tool support. Outgoing from the sequential legacy software, a very precise executable model (using the ABS language) is extracted. The model is then analyzed and parallelized. From the parallel model executable and optimized code is generated. The advantage is that the model language ABS has been developed to be analyzable. This allows one to apply a wide range of far reaching analyses tools to achieve
- correctness proofs,
- scalability and
- a high degree of automation.
The described workflow is illustrated in the figure below: