GROMACS 2024.4 release notes#
This version was released on October 31st, 2024. These release notes document the changes that have taken place in GROMACS since the previous 2024.3 version, to fix known issues. It also incorporates all fixes made in version 2023.5 and earlier, which you can find described in the Release notes.
Fixes where mdrun could behave incorrectly#
Fix missing non-bonded interactions close to cut-off with GPUs#
The GPU rolling pruning employed with the dual pair list setup used
an incorrect list part calculation. This caused a few parlist entries
not to be updated after the initial pruning, leading to a small number
of missing pair interactions. Unless the pair list lifetime nstlist
was very large, these interactions were close to the cut-off and therefore
the errors were small.
Affected simulations: all GPU accelerated runs using GROMACS versions
starting with the 2018 release with dual pair list enabled.
This is enabled automatically with default Verlet buffering, but
it is disabled if the Verlet buffer is manually set
(or if the GMX_DISABLE_DYNAMICPRUNING environment variable is set).
Impact: Our analysis shows that the error caused by the missing interactions
does not have a measurable effect on either forces or energy conservation
in most atomistic simulations with default simulation settings.
While a manually increased nstlist, hence increased (outer) pair list lifetime,
leads to an increase in the number of missing interactions, for useful value of
nstlist the effect is still negligible.
There can be measurable effects in systems with vacuum, gas, or with
excessively large nstlist values. In our testing we could not detect artifacts
apart from systems crashing, so the chance of undetected incorrect results is small.
Add effect of perturbed masses to foreign Hamiltonian differences#
The effect of perturbed masses was missing from the foreign Hamiltonian
differences. This meant that effects of perturbed masses were ignored
by gmx bar. Note that perturbed masses did contribute to dH/dlambda.
Fix illegal memory access with more VCM than T-coupling groups#
When there were more center of mass motion removal groups than temperature coupling groups, mdrun could crash due to illegal memory access. This bug did not affect the simulation results.
Fix crashes when some atoms are not part of a VCM group#
Uninitialized memory was read when some atoms were not part of a center of mass motion removal group. Often this memory contained zero and everything would work as it should. But when this memory was not zero, the simulation would crash within a few steps.
Fix parameter handling for coarse-grained bonded potentials#
Free-energy perturbation has never been supported for restraint angle,
restraint dihedral, and combined bending-torsion potentials sometimes
used in coarse-grained force fields. This led to the hypothetical
B-state parameters being incorrectly handled in gmx grompp and
then incorrectly serialized to/from the .tpr file.
This was generally benign, even when such functional forms were in use. However, when other interactions were perturbed, the incorrect parameter handling made it look like the interactions with wrongly implemented parameters were intended to be perturbed. This condition was sometimes achieved when using alchemical-embedding system-preparation workflows.
Now the hypothetical B-state parameters are handled correctly, so FEP-based workflows can succeed. Perturbation of the above-mentioned interaction types is still not supported, however.
Fix Colvars output files always written to the working directory.#
Colvars output files are now written in the same folder as edr file.
Forbid the usage of triangle constraints with -update gpu#
Prevent using update on GPU if there are triangle constraints.
gmx_mpi mdrun could hang when using GPUs and separate PME ranks#
A logic error in task assignment for -nb gpu -pme cpu and separate PME ranks
made the (default) -ddorder interleave hang. Now it works.
Dynamic load balancing was ineffective when special forces were present#
The timing for DLB included the calculation of special forces, e.g. pull code and rotation. As these require communication, imbalance might not have been measured. Now special forces are excluded from the timings.
Fix incorrect memory access with perturbed non-bonded and OpenMP#
Old, invalid indices could be flagged for reduction of force buffers over OpenMP threads for perturbed non-bonded interactions. This could never lead to incorrect results or crashes. But performance might be improved now.
Fixes for gmx tools#
grompp checked incorrect B-state charges with free-energy decoupling#
When the couple-moltype options was used, gmx grompp would check the
A-state charges instead of the B-state charges. This would lead to incorrect
or no warnings when B-state of the system had non-zero net charge.
grompp, and mdrun, could exit with large mass differences#
When masses of atoms differed more than a factor 327, both gmx grompp
and gmx mdrun could exit with an assertion failure about an infinite
enery drift estimate.
Fix dump crash with Colvars values#
Fix gmx dump crash when trying to output the binary Colvars state file.
Fix element and atom number deduction in during preprocessing#
Two-letter atom names like BR or CL were not correctly handled in the preprocessing
during gmx editconf, gmx pdb2gmx and gmx grompp, leading to incorrect element
and atom number assignment. This could lead to incorrect element names in output files
and possibly incorrect behaviour in QMMM simulations.
Fix hang observed with NVSHMEM enabled PME-PP force transfers#
A hang in PME-PP force transfers during NVSHMEM runs was resolved, occurring in certain conditions. It should also be noted that the NVSHMEM enabled PME-PP force transfers does not support charge perturbation.
Fixes that affect portability#
Fix physical validation with Pymbar version 4#
Pymbar version 4 has a different API compared to version 3. Now we support using either of those versions and internally handle the API differences.
Fix compatibility with VkFFT 1.3.5#
Fix crash on GROMACS shutdown when external VkFFT 1.3.5 is used.
Miscellaneous#
Work around FetchContent warnings in CMake 3.30 and newer#
CMake 3.30 began to complain about GROMACS use of FetchContent, so now we tell such new versions to use the old policy.
Fix various crashes when GMX_USE_COLVARS=NONE#
Colvars MDModule did not defined the Colvars custom mdp variables when the Colvars library was not compiled, preventing tools read correctly a tpr file generated with a GROMACS-Colvars version. Prevent also the creation of a tpr file if Colvars module is activated whereas GROMACS was not compiled with Colvars. Proper exit if a Colvars simulation is launched whereas GROMACS was not compiled with Colvars.
Fix reading cgroups in some kubernetes containers#
Modern versions of kubernetes/docker do not appear to mount /etc/mtab in the containers, and if we did not find cgroups we would not detect CPU limits set through cgroups. Fixed by reading /proc/mounts instead. This will only influence performance when running in (some) containers.
Collected fixes in the Colvars library#
Several bugs, both recent and long-standing, have recently been fixed in the Colvars library: in the list below, the links corresponds to issues or pull request in the Colvars repository.
Fixed undefined behavior when getting the current working directory from
std::filesystem, which could have affected multiple-walkers metadynamics runs (Colvars PR 728).Fixed the gradients and the metric functions of collective variables of
distanceDirtype (Colvars PR 724).Fixed the definition of an
orientationtype collective variable in a rotated frame of reference (Colvars PR 715).Implemented the contribution of fitting to the forces applied onto of variables with vector values defined in a rotated fram (Colvars PR 713).
Fixed a crash in certain metadynamics simulations without using interpolating grids (Colvars PR 706).
More consistent behavior when defining multiple biases and running with more than one thread per task (Colvars PR 694).
Prevented the creation of spurious output files for runtime histograms (Colvars PR 675).
Enable NVCC flags checks for Windows#
When building GROMACS for Windows with CUDA support, the checks for testing compatible compute architectures were disabled. Hence, GROMACS was trying to compile for all of them which can induced failed builds when CUDA is not compatible anymore with old compute architectures.