You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Task: Evaluate the performance of cuSOLVER and PardisoMKL linear solvers vs Basker/KLU2 for our problems. If they are faster, add a switch in solver creation to use them when availible.
Why: I recently saw some papers talking about the performance of these "native" solvers. Since the linear solve step is so important to our performance, it is worth evaluating them for our problem. Since we will be using them through the Amesos2 interface, the impact on our code will be minimal. These solvers are chosen because they are likely to be present on many target machines and well optimized for the single-node case.
Done means: cuSolver and PardisoMKL have been evaluated. If they are faster, a switch to use them is added to the Solver constructor. Default use of Basker/KLU2 is maintained when those TPLs are not available.
The text was updated successfully, but these errors were encountered:
Task: Evaluate the performance of cuSOLVER and PardisoMKL linear solvers vs Basker/KLU2 for our problems. If they are faster, add a switch in solver creation to use them when availible.
Why: I recently saw some papers talking about the performance of these "native" solvers. Since the linear solve step is so important to our performance, it is worth evaluating them for our problem. Since we will be using them through the Amesos2 interface, the impact on our code will be minimal. These solvers are chosen because they are likely to be present on many target machines and well optimized for the single-node case.
Done means: cuSolver and PardisoMKL have been evaluated. If they are faster, a switch to use them is added to the Solver constructor. Default use of Basker/KLU2 is maintained when those TPLs are not available.
The text was updated successfully, but these errors were encountered: