west6.2 test-suite issue

[doronwu]

Hi all,

Recently I tried to compile the latest version of west-code on NERSC. I fellow the instruction on the website and compile it with two step (as shown in compile.txt)

Compile.txt

The issue is when I tried to do test-suites, all the QE calculation passed the test, but the majority of west test are failed (in test-report.log).

test-report.log

I also tried to do 6.0v on NERSC with the same version of python and fftw on the 6.0v website but the same issue occured. I wonder if anyone has the same issue?

[vyu16]

@doronwu I think the failures you encountered were due to using 8 MPI processes when running pw.x. You can rerun the tests with export PARA_PREFIX_QE=srun -n 2 in the file test-suite/maketest.inc, and you should see very small errors.

The reference numbers of the test suite were obtained by running pw.x with 2 processes. Using a different number of processes leads to slightly different Kohn-Sham wave functions, which in turn lead to different results in WEST. Typically the difference in the quasiparticle energy is below 1 meV, as one would expect. However, in test009 and test014, some bands have multiple roots of the quasiparticle equations. Therefore, when small differences in the starting wave functions are present, the iterative solution of the quasiparticle equations can converge to different solutions, causing seemingly huge errors in the tests.

For more information about the issue of multiple roots, see, for example, section 8 of https://doi.org/10.1021/acs.jctc.6b00114. A tutorial on this topic will be available in the next release of WEST.

To make the test suite more robust, we will think about replacing the numerically unstable test cases with more stable ones.

Thanks,
Victor