MueLu: uncoupled aggregation kokkos refactor
Created by: lucbv
@trilinos/muelu
Expectations
Aggregation algorithm need to be refactored to run efficiently on new architectures
Current Behavior
aggregation runs on new platform but large parts of the code are still serial
Motivation and Context
This will be useful for simulation on advanced platform coming in production.
Definition of Done
-
rewrite aggregation algorithms using kokkos -
check that the new algorithms compile with both OpenMP and CUDA nodes -
verify that the aggregates created using these algorithm are reasonable for MueLu's use
Possible Solution
pull request xxx is submitted with a proposed implementation that should scale reasonable well at least with OpenMP.