Open lorisercole opened 10 months ago
@xubo-wang should know about the ghf test, it has been failing for a bit I think.
About the number of tasks, this is because the convention in our code is to increase the number of samples with an increasing number of tasks. So the sampling input options are per task, and the tests only work with four tasks.
Default tests
I have run the tests on 4 MPI tasks (as hardcoded in the tests scripts), and they all pass flawlessly, except for the
DQMC/multislater_ghf_gi
one, for which I get:Different number of MPI tasks
When I tried to do the same on a different number of MPI tasks, for example 2, 6, 8, 16, most (or all) tests fail. Often the energy difference with respect to the reference value can be of the order of 0.1 or 0.01, other times it is of the order of 0.001, so well above the set tolerances (1e-6 or 1e-7).
Example test output
I have tried building Dice with GCC 13.2, GCC 11.4, and ICC 2021.10, and I always get these inconsistencies. I am linking it with boost@1.82, hdf5@1.14.1, MKL 2023.2, and OpenMPI.
Have you ever seen this behavior and do you understand where these differences may come from? Or is within the expected statistical fluctuations due to the stochastic nature of the method? Thanks