2 =======================
3 This section discusses features in :ref:`gmx mdrun` that don't fit well
6 .. _single-point energy:
8 Re-running a simulation
9 -----------------------
10 The rerun feature allows you to take any trajectory file ``traj.trr``
11 and compute quantities based upon the coordinates in that file using
12 the model physics supplied in the ``topol.tpr`` file. It can be used
13 with command lines like ``mdrun -s topol -rerun traj.trr``. That :ref:`tpr`
14 could be different from the one that generated the trajectory. This
15 can be used to compute the energy or forces for exactly the
16 coordinates supplied as input, or to extract quantities based on
17 subsets of the molecular system (see :ref:`gmx convert-tpr` and
18 :ref:`gmx trjconv`). It is easier to do a correct "single-point" energy
19 evaluation with this feature than a 0-step simulation.
21 Neighbor searching is performed for every frame in the trajectory
22 independently of the value in :mdp:`nstlist`, since
23 :ref:`gmx mdrun` can no longer assume anything about how the
24 structures were generated. Naturally, no update or constraint
25 algorithms are ever used.
27 The rerun feature cannot, in general, compute many of the quantities
28 reported during full simulations. It does only take positions as input
29 (ignoring potentially present velocities), and does only report potential
30 energies, volume and density, dH/dl terms, and restraint information.
31 It does notably not report kinetic, total or conserved energy, temperature,
34 Running a simulation in reproducible mode
35 -----------------------------------------
36 It is generally difficult to run an efficient parallel MD simulation
37 that is based primarily on floating-point arithmetic and is fully
38 reproducible. By default, :ref:`gmx mdrun` will observe how things are going
39 and vary how the simulation is conducted in order to optimize
40 throughput. However, there is a "reproducible mode" available with
41 ``mdrun -reprod`` that will systematically eliminate all sources of
42 variation within that run; repeated invocations on the same input and
43 hardware will be binary identical. However, running in this mode on
44 different hardware, or with a different compiler, etc. will not be
45 reproducible. This should normally only be used when investigating
48 Halting running simulations
49 ---------------------------
51 When :ref:`gmx mdrun` receives a TERM or INT signal (e.g. when ctrl+C is
52 pressed), it will stop at the next neighbor search step or at the
53 second global communication step, whichever happens later.
54 When :ref:`gmx mdrun` receives a second TERM or INT signal and
55 reproducibility is not requested, it will stop at the first global
57 In both cases all the usual output will be written to file and
58 a checkpoint file is written at the last step.
59 When :ref:`gmx mdrun` receives an ABRT signal or the third TERM or INT signal,
60 it will abort directly without writing a new checkpoint file.
61 When running with MPI, a signal to one of the :ref:`gmx mdrun` ranks
62 is sufficient, this signal should not be sent to mpirun or
63 the :ref:`gmx mdrun` process that is the parent of the others.
65 Running multi-simulations
66 -------------------------
67 There are numerous situations where running a related set of
68 simulations within the same invocation of mdrun are necessary or
69 useful. Running a replica-exchange simulation requires it, as do
70 simulations using ensemble-based distance or orientation restraints.
71 Running a related series of lambda points for a free-energy
72 computation is also convenient to do this way.
75 :ref:`configuring |Gromacs| with an external MPI library <mpi-support>`
77 simulations can communicate. The ``n`` simulations within the set can
78 use internal MPI parallelism also, so that ``mpirun -np x mdrun_mpi``
79 for ``x`` a multiple of ``n`` will use ``x/n`` ranks per simulation.
81 There are two ways of organizing files when running such
82 simulations. All of the normal mechanisms work in either case,
83 including ``-deffnm``.
86 You must create a set of ``n`` directories for the ``n`` simulations,
87 place all the relevant input files in those directories (e.g. named
88 ``topol.tpr``), and run with
89 ``mpirun -np x gmx_mpi mdrun -s topol -multidir <names-of-directories>``.
90 If the order of the simulations
91 within the multi-simulation is significant, then you are responsible
92 for ordering their names when you provide them to ``-multidir``. Be
93 careful with shells that do filename globbing dictionary-style, e.g.
94 ``dir1 dir10 dir11 ... dir2 ...``. This option is generally the
95 most convenient to use. ``gmx mdrun -table`` for the group cutoff-scheme
96 works only in this mode.
98 Examples running multi-simulations
99 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
103 mpirun -np 32 gmx_mpi mdrun -multidir a b c d
105 Starts a multi-simulation on 32 ranks with 4 simulations. The input
106 and output files are found in directories ``a``, ``b``, ``c``, and ``d``.
110 mpirun -np 32 gmx_mpi mdrun -multidir a b c d -gputasks 0000000011111111
112 Starts the same multi-simulation as before. On a machine with two
113 physical nodes and two GPUs per node, there will be 16 MPI ranks per
114 node, and 8 MPI ranks per simulation. The 16 MPI ranks doing PP work
115 on a node are mapped to the GPUs with IDs 0 and 1, even though they
116 come from more than one simulation. They are mapped in the order
117 indicated, so that the PP ranks from each simulation use a single
118 GPU. However, the order ``0101010101010101`` could run faster.
120 Running replica-exchange simulations
121 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
123 When running a multi-simulation, using ``gmx mdrun -replex n`` means that a
124 replica exchange is attempted every given number of steps. The number
125 of replicas is set with ``-multidir`` option, described
126 above. All run input files should use a different value for the
127 coupling parameter (e.g. temperature), which ascends over the set of
128 input files. The random seed for replica exchange is set with
129 ``-reseed``. After every exchange, the velocities are scaled and
130 neighbor searching is performed. See the Reference Manual for more
131 details on how replica exchange functions in |Gromacs|.
133 Controlling the length of the simulation
134 ----------------------------------------
136 Normally, the length of an MD simulation is best managed through the
137 :ref:`mdp` option :mdp:`nsteps`, however there are situations where
138 more control is useful. :samp:`gmx mdrun -nsteps 100` overrides the :ref:`mdp`
139 file and executes 100 steps. :samp:`gmx mdrun -maxh 2.5` will terminate the
140 simulation shortly before 2.5 hours elapse, which can be useful when
141 running under cluster queues (as long as the queuing system does not
142 ever suspend the simulation).