All,
Our proposal to participate in the Argonne GPU hackathon event [1] was
approved last week. The MOAB development team would like to solicit
ideas for mini-apps that would have an impact on the library usage in
hybrid-GPU architectures.
We are currently planning to use the following mini-applications in
order to influence the future design of the library to support solvers
and multiphysics applications.
1) A fast hybrid-GPU friendly FEM assembly test app based on MOAB
representation hierarchy of material/type/element/quadrature traversal
in 2d and 3d for solving the Poisson problem. This mini-app would help
determine performance characteristics of the mesh traversal and
optimizations needed to provide better local ordering (for mixed
meshes), element coloring with Metis or Zoltan (to reduce contention)
and exploration of support for linear algebra packages like Eigen3 [2]
or Gingko [3] for operator assembly and solution computation.
2) Optimizing the underlying algorithms for unstructured mesh
intersection computations using advancing front and/or Kd-tree based
implementations for Climate applications. Lessons learned will also
directly impact the mbcoupler library to help provide solution
transfer mechanisms for a wide range of multiphysics applications.
We also realize there are many other potential mini-app ideas to
address such as exploration of better polygon/polyhedra datastructure
representations, efficient ray tracing on GPUs, etc to name a few. And
so we wanted to open up the discussion to better understand what would
provide the best impact for your work, and which areas of the library
to prioritize for hybrid-GPU architectures. All comments welcome.
Best,
Vijay
[1] https://www.gpuhackathons.org/index.php/event/argonne-gpu-hackathon-2021
[2] https://eigen.tuxfamily.org/index.php?title=Main_Page
[3] https://ginkgo-project.github.io/
