Scope out move to Tempus #11
Comments
|
@ikalash We are currently working on this. I'm reviving this so that we can track it. |
|
Thanks for reviving the issue, @lxmota! To give everyone a status update, I have hooked up Albany to Tempus via the Piro::TempusSolver class, which I added to the Piro package of Trilinos. There are several tests in Albany demonstrating syntax for how to call Tempus through Albany (one can see which ones by doing grep -r Tempus on the examples directory in Albany). Note that Tempus does not yet have schemes for 2nd order PDEs like LCM problems yet. @lxmota and I are working on adding these schemes. |
|
Today, I learned that Tempus has been granted copyright assertion. It will be moving to Trilinos on 2/14/17. |
|
It looks like Tempus just landed in the Trilinos https://github.com/trilinos/Trilinos/tree/develop/packages/tempus |
|
To give an update towards this issue: I have added the Newmark Beta scheme (and more general HHT Alpha scheme) to Tempus and verified its convergence and properties. Attached are some results. These were generated outside of Albany and are running as Tempus tests. I am currently working on integrating the new scheme into Albany. |
|
The new Newmark Beta scheme in Tempus has been hooked up to Albany and appears to be behaving as expected on a simple problem involving the rotation of a cube. Please see commits c151484, af056a5, 42f94c8 and db462ad. An important detail to get the correct behavior was to switch from a Linear Elastic material model to a Neohookean (nonlinear) material model, as discussed during this week's Albany concall. This is because the latter material model is not implemented in a way to enable large rigid body motions / rotations. With a Linear Elastic material models, the cubes expand in a non-physical way as they rotate. |
Adding input files to run Newmark Beta scheme through Piro and Tempus in explicit mode (beta = 0.0, gamma = 0.5). Studies of this scheme's performance are still underway, but results are promising. For single cube geometry, my "back-of-the-envelope" calculations say that the stable time-step is dt <= 7.211e-5; scheme is stable for dt = 7e-5, unstable for dt = 8e-5. For larger cube geometry, my "back-of-the-envelope" calculations say that the stable time-step is dt <= 1.89e-5; scheme is stable for dt = 1e-5, unstable for dt = 2e-5. Some caveats: - The essential BCs in Albany are not implemented in a way that will be handled correctly with an explicit time-integration scheme at the present time (see issue #13). This problem works explicitly b/c there are no essential BCs in the problem specification. Do not try explicit Newmark Beta schemes w/ essential BCs, as this will not work correctly in general. - The explicit Newmark Beta scheme in Piro and Tempus has not been optimized to do mass-lumping and a trivial inversion of the resulting lumped (diagonal) mass matrix. Since inversions of the mass matrix are happening under the hood when running explicit, please do not expect explicit to be faster than implicit at the present time. The optimizations involving lumping can be done in a straightforward way, once the scheme w/o lumping is tested more thoroughly.
Not hooked up as a test yet until Tempus is verified more thoroughly.
…tion. Not hooked up as a test yet until Tempus is verified more thoroughly.
Tempus verification. (pair program with Irina)
…tion for Tempus (pair program with Irina).
|
Tempus is being used consistently now, so closing this issue. |
Tempus is a new time integration package in Trilinos.
The text was updated successfully, but these errors were encountered: