Bug #2876
2019.1 make check fails on AVX and AVX2 (Intel 2018u3)
Description
Hello
make check fails in 2019.1 on two different platforms. In both cases I used Intel 2018 update 3 compilers.
- on AVX, 3 SIMD tests fail, but only in single precision
- on AVX2, 1 MdrunNonIntegratorTests test (and consequently MdrunMpiTests) fails, but only in double precision
I attach for both cases the log file from the error, as well as the output from gmx* -version
FYI: I used the same Intel compiler with GROMACS 2018.3 without problem. On AVX, I had to add "-no-ftz" as a compiler flag to pass the tests.
Kind regards
Stefan
Related issues
History
#1 Updated by Erik Lindahl about 2 years ago
These look very benign, so it's likely just a matter of allowing a bit more freedom in the tests for the Intel compiler.
#2 Updated by Erik Lindahl about 2 years ago
- Assignee set to Erik Lindahl
#3 Updated by Stefan Becuwe about 2 years ago
- File regression-avx-2019.1-intel-2019u2.txt regression-avx-2019.1-intel-2019u2.txt added
- File regression-avx2-2019.1-intel-2019u2.txt regression-avx2-2019.1-intel-2019u2.txt added
Hello Erik
Update: I also tried the Intel 2019 update 2 compilers. Adding the '-no-ftz' flag on the AVX machine, all tests pass on both platforms!
I also tried the regression tests with Intel 2019 update 2. I found one case that goes wrong on both AVX and AVX2 in both single and double precision, and one extra case that fails on AVX2. See attachment for used flags and errors.
Kind regards
Stefan
#4 Updated by Roland Schulz almost 2 years ago
The freeenergy error with ICC19u2 with AVX2 I can't reproduce. Could you please attach the mdrun.out, md.log files for those failing freeenergy tests?
The essentialdynamics error I can reproduce. Thanks for the report. The error is in the testing harness not in GROMACS. So you don't have to worry about the error. I crated #2883 for this bug.
#5 Updated by Roland Schulz almost 2 years ago
- Related to Bug #2883: essentialdynamics fail with Intel MPI added
#6 Updated by Stefan Becuwe almost 2 years ago
- File freeenergy.tgz freeenergy.tgz added
- File gmx-avx2-interactive-logfile gmx-avx2-interactive-logfile added
Hello Roland
It seems those two specific tests (expanded and simtemp) from freeenergy are influenced by the environment in one way or another.
If I login directly to a compute node or login node, all regression tests pass.
If I start an interactive job (qsub -I -lnodes=1:ppn=28 (all cores)) and run the regression tests, those two tests fail. I reproduced this situation twice.
Let's hope this is useful information.
Kind regards
Stefan