Belos: PseudoBlockGmres complex tests failing
Created by: mhoemmen
@trilinos/belos
Belos_bl_gmres_complex_hb_3 and Belos_bl_gmres_complex_hb_4 fail. See e.g., the perseus test on the Dashboard. They fail for me in both MPI debug and non-MPI release builds. The tests in question do NOT exercise either Epetra or Tpetra; they use Belos' custom linear algebra implementation that Belos has long used for testing. I'm not sure how long the tests have been failing. I played with the test executable and matrix, and did get them to converge when I changed the restart length, e.g.,
/Belos_bl_gmres_complex_hb.exe --verbose --filename=mhd1280a.cua --pseudo --num-restarts=1 --frequency=1 --subspace-length=800 --num-restarts=1000
Thus, I'm not sure what's going on here. Could it be that the expected convergence rate depends on a particular random right-hand side, and I'm using a different system library's pseudorandom number generator? (I'm using Clang 3.7 on Mac.)