search

mdolab / adflow

ADflow is a finite volume RANS solver tailored for gradient-based aerodynamic design optimization.
Other
220 stars 98 forks source link

AMG regression test fails with three levels #324

Open eirikurj opened 11 months ago

eirikurj commented 11 months ago

Description

Nightly docker integration tests on the u22-gcc-ompi-stable x86 image is failing on the AMG test (see here). This seems to be a result of a newly merged PR https://github.com/mdolab/adflow/pull/235.

The test_amg.py:TestAMGReal_2_different_levels_and_iterations.test_convergence AMG regression test is failing to converge.

Steps to reproduce issue

This can be reproduced with the failed image pushed to DockerHub, i.e., mdolab/public:u22-gcc-ompi-stable-amd64-failed (Digest: sha256:76e6eb4be1c823744ec7cfc46fa38e5edc9897b8cf9d54dc08d208cd5db1d2f2). Using the image, download the input test files, and run only the AMG regression tests testflo -s -v -n 1 test_amg.py.

Using the previous commit git checkout 559a1deee07126c521e7a2e18f3968dd0ef7b061, rebuilding and pip installing, then all AMG tests do pass successfully.

Current behavior

Currently, the case does not converge and test fails.

Expected behavior

Case should converge and test pass.

Code versions

Docker image mdolab/public:u22-gcc-ompi-stable-amd64-failed downloaded today (Digest: sha256:76e6eb4be1c823744ec7cfc46fa38e5edc9897b8cf9d54dc08d208cd5db1d2f2)

sseraj commented 7 months ago

This was fixed in #325 by changing the test. The root cause is unclear except that using too many AMG levels potentially causes memory-related issues.

eirikurj commented 7 months ago

@sseraj, if there is a known issue, I would prefer to either document it somewhere in the docs or create a separate issue.

sseraj commented 7 months ago

This issue has all the relevant information, so maybe it is best to just reopen this