This repository has been archived by the owner on Jan 4, 2022. It is now read-only.
ECP 6: Establish and improve strong scaling of linear solves on high aspect ratio meshes #60
Labels
Comments
|
@spdomin I'd like to start by identifying a set of canonical tests with high aspect ratios and benchmarking their performance. Ideally, the tests wouldn't be too large, so that some analysis might be possible. |
|
Sounds great. The heatedWaterChannel is one that you used before. I can find more next week. |
crtrott
pushed a commit
to crtrott/Nalu
that referenced
this issue
Feb 6, 2017
…in#60) The version of the element-algorithm that uses std::vectors for scratch arrays (TestElemAlgorithmWithVectors) can not execute correctly in true multi-threaded mode since the vectors can't feasibly be allocated on a per-thread basis. Thus, this version of the element-algorithm is now restricted to explicitly *not* run multi-threaded. The version of the element-algorithm that uses Kokkos::views for scratch arrays (TestElemAlgorithmWithViews) now executes correctly in true multi-threaded mode (OpenMP). It uses the SharedMemView construct that Christian had prototyped. Unfortunately this means that it can not use the loop-encapsulation mechanism outlined in recent slides. A new version of the element-algorithm has now been added (TestElementAlgorithmWithTemplate) which uses a templated kernel function for the inner-loop-body, meaning that scratch arrays are automatic arrays allocated at compile time. Thus no resize operations are needed. This works in multi-threaded mode, and is the fastest of the 3 approaches, by a small margin.
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Activities:
The text was updated successfully, but these errors were encountered: