guest ::
| Home > Publications database > From Iteration Counting to Applications with pySDC |
| Home > Publications database > From Iteration Counting to Applications with pySDC |
| Conference Presentation (After Call) | FZJ-2024-04851 |
;
2024
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2024-04851
Abstract: Many parallel-in-time (PinT) algorithms replace the serial and, in this regard, direct way of time stepping by algorithms that iterate on multiple time steps in parallel.Efficient parallel implementation is a demanding exercise and may exceed the scope of more math-based PinT projects.Instead, PinT research typically resorts to "counting iterations" of the algorithm as a means of measuring its performance independently of its actual implementation.However, this hides many non-negligible sources of computational cost, such as communication.If the true parallel efficiency of candidate algorithms is under consideration, this needs to be obtained in a more coherent way.We present here one prototyping library which aims to cover both aspects of PinT research: method development and fair efficiency testing.pySDC is a Python code that, while not providing production-level performance, allows users to detect pitfalls in parallel algorithms before committing to optimized implementations or rigorous mathematical analysis.Parallel efficiency can be estimated by comparing to various time-serial algorithms implemented in the same framework.The modular structure allows users to easily apply a new algorithm to a wide range of problems and configurations without awareness of all the intricacies of the code.pySDC is publicly hosted on GitHub and well tested and documented in order to provide new users with a smooth and rapid start.
|
The record appears in these collections: |
PDF