Closed maxrudolph closed 8 years ago
I have talked to Louise about this, and she will forward the problem to Eric, who has the most experience on the tracer code, so maybe there will be some improvement in the near future.
At the very least we may want to remove the assertion that the particle count never changes. I'm not sure it is realistic to expect that we never lose a tracer, especially where velocity gradients are sharp.
This should finally be closed with #411. @maxrudolph do you want to check? I never encountered these problems during my recent tests on more than a thousand cores. In the new version particles are allowed to get lost, although it almost never happens. The number of particles is tracked in the statistics file.
@gassmoeller Using the development versions of aspect and deal.ii, it appears that this problem is solved! Thanks for contributing your vastly improved tracer code!
I've encountered crashes related to using tracers and MPI with two or more processes. Attached, you'll find a simple parameter file that will produce an error after several hundred (800-900) timesteps when run with two MPI processes. This occurred on ubuntu 14.04 and also on OS X. I compiled deal.ii and ASPECT with clang 3.4-lubuntu3. I'm using openmpi bundled with ubuntu (which uses gcc 4.8.2 by default but I've used the OMPI_CXX environment variable to make the wrappers use clang because compiling deal.ii results in internal compiler errors with gcc).
------- Input file: