Test failure in GRHydro's balsara4_1d

Create issue
Issue #1730 closed
Erik Schnetter created an issue

I see test failures in GRHydro's balsara4_1d. The maximum absolute difference is about 1e-10, the maximum relative difference about 1e-11. This seems like a problem with tolerances. Can someone with GRHydro experience check?

Keyword:

Comments (9)

  1. Roland Haas
    • removed comment

    This is most likely a tolerance issue. Balsara4 uses a very high Lorentz factor and we already had to increase the tolerance for the 2d version of the test to 1e-8 for Bcom in 4c0d195. The relevant part of test.ccl would be:

    TEST balsara4_1d
    {
            NPROCS 2
            RELTOL 1e-8 grhydro::bcom
            RELTOL 1e-9 hydrobase::w_lorentz
            RELTOL 1e-10 hydrobase::(eps|rho)
    }
    

    Which machine/compiler/options did you use to see the test failure?

  2. Erik Schnetter reporter
    • removed comment

    This is e.g. OS X or Ubuntu, with gfortran 4.9.2, without optimization.

  3. Roland Haas
    • removed comment

    Hmm. This is new then. I use gfortran 4.9.2 on my (Debian) workstation and the test passes there (kind of unsurprising since the data was generated on that machine) and I also tested it on a Ubuntu VM shortly after the release. I'll check once more if it still passes or fails and if it fails, when it started to fail.

  4. Roland Haas
    • removed comment

    Replying to [comment:3 rhaas]:

    Hmm. This is new then. I use gfortran 4.9.2 on my (Debian) workstation and the test passes there (kind of unsurprising since the data was generated on that machine) and I also tested it on a Ubuntu VM shortly after the release. I'll check once more if it still passes or fails and if it fails, when it started to fail.

    Addendum: I tested with optimization though. Unless we are using fast-math with gfortran the results should not differ, but then, Fortran does not really guarantee any particular ordering for floating point evaluations. I'll test with and without optimizations.

  5. Erik Schnetter reporter
    • removed comment

    I see this error report:

       grhydro-divb.x.asc: differences below tolerance on 10 lines
       hydrobase-bvec.x.asc: differences below tolerance on 2 lines
       hydrobase-eps.x.asc: differences below tolerance on 30 lines
       hydrobase-press.x.asc: differences below tolerance on 30 lines
       hydrobase-rho.x.asc: substantial differences
          significant differences on 28 (out of 30) lines
          maximum absolute difference in column 13 is 1.15789600130256e-11
          maximum relative difference in column 13 is 1.1578960013032e-11
          (insignificant differences on 2 lines)
       hydrobase-vel.x.asc: differences below tolerance on 30 lines
       hydrobase-w_lorentz.x.asc: substantial differences
          significant differences on 28 (out of 30) lines
          maximum absolute difference in column 13 is 2.58999932611914e-10
          maximum relative difference in column 13 is 1.15799330402901e-11
          (insignificant differences on 2 lines)
    

    Yes, I think these tolerances would avoid this error.

  6. Roland Haas
    • changed status to resolved
    • removed comment

    Those tolerances were already present however the patterns in test.ccl used the old filenames hydrobase::rho instead of hydrobase-rho. I have updated test.ccl in EinsteinEvolve hash c23d7d8 .

  7. Log in to comment