| blob | 320eea5929b5f7934ab28c84dd19186ecf802465 |
1 .. _managing long simulations:
3 Managing long simulations
4 =========================
6 Molecular simulations often extend beyond the lifetime of a single
7 UNIX command-line process. It is useful to be able to stop and
8 restart the simulation in a
9 way that is equivalent to a single run. When :ref:`gmx mdrun` is
10 halted, it writes a checkpoint file that can restart the simulation
11 exactly as if there was no interruption. To do this, the checkpoint
12 retains a full-precision version of the positions and velocities,
13 along with state information necessary to restart algorithms e.g.
14 that implement coupling to external thermal reservoirs. A restart can
15 be attempted using e.g. a :ref:`gro` file with velocities, but since
16 the :ref:`gro` file has significantly less precision, and none of
17 the coupling algorithms will have their state carried over, such
18 a restart is less continuous than a normal MD step.
20 Such a checkpoint file is also written periodically by :ref:`gmx
21 mdrun` during the run. The interval is given by the ``-cpt`` flag to
22 :ref:`gmx mdrun`. When :ref:`gmx mdrun` attemps to write each
23 successive checkpoint file, it first renames the old file with the
24 suffix ``_prev``, so that even if something goes wrong while writing
25 the new checkpoint file, only recent progress can be lost.
27 :ref:`gmx mdrun` can be halted in several ways:
29 * the number of simulation :mdp:`nsteps` can expire
30 * the user issues a termination signal (e.g. with Ctrl-C on the terminal)
31 * the job scheduler issues a termination signal when time expires
32 * when :ref:`gmx mdrun` detects that the length specified with
33 ``-maxh`` has elapsed (this option is useful to help cooperate with
34 a job scheduler, but can be problematic if jobs can be suspended)
35 * some kind of catastrophic failure, such as loss of power, or a
36 disk filling up, or a network failing
38 To use the checkpoint file for a restart, use a command line such as
40 ::
42 gmx mdrun -cpi state
44 which directs mdrun to use the checkpoint file (which is named
45 ``state.cpt`` by default). You can choose to give the output
46 checkpoint file a different name with the ``-cpo`` flag, but if so
47 then you must provide that name as input to ``-cpi`` when you later
48 use that file. You can
49 query the contents of checkpoint files with :ref:`gmx check` and
50 :ref:`gmx dump`.
52 Appending to output files
53 -------------------------
55 By default, :ref:`gmx mdrun` will append to the old output files. If
56 the previous part ended in a regular way, then the performance data at
57 the end of the log file will will be removed, some new information
58 about the run context written, and the simulation will proceed. Otherwise,
59 mdrun will truncate all the output files back to the time of the last
60 written checkpoint file, and continue from there, as if the simulation
61 stopped at that checkpoint in a regular way.
63 You can choose not to append the output files by using the
64 ``-noappend`` flag, which forces mdrun to write each output to a
65 separate file, whose name includes a ".partXXXX" string to describe
66 which simulation part is contained in this file. This numbering starts
67 from zero and increases monotonically as simulations are restarted,
68 but does not reflect the number of simulation steps in each part. The
69 :mdp:`simulation-part` option can be used to set this number manually
70 in :ref:`gmx grompp`, which can be useful if data has been lost,
71 e.g. through filesystem failure or user error.
73 Appending will not work if any output files have been modified or
74 removed after mdrun wrote them, because the checkpoint file maintains
75 a checksum of each file that it will verify before it writes to them
76 again. In such cases, you must either restore the file, name them
77 as the checkpoint file expects, or continue with ``-noappend``. If
78 your original run used ``-deffnm``, and you want appending, then
79 your continuations must also use ``-deffnm``.
81 Backing up your files
82 ---------------------
84 You should arrange to back up your simulation files frequently. Network
85 file systems on clusters can be configured in more or less conservative
86 ways, and this can lead :ref:`gmx mdrun` to be told that a checkpoint
87 file has been written to disk when actually it is still in memory
88 somewhere and vulnerable to a power failure or disk that fills or
89 fails in the meantime. The UNIX tool rsync can be a useful way to
90 periodically copy your simulation output to a remote storage location,
91 which works safely even while the simulation is underway. Keeping a copy
92 of the final checkpoint file from each part of a job submitted to a
93 cluster can be useful if a file system is unreliable.
95 Extending a .tpr file
96 ---------------------
98 If the simulation described by :ref:`tpr` file has completed and should
99 be extended, use the :ref:`gmx convert-tpr` tool to extend the run, e.g.
101 ::
103 gmx convert-tpr -s previous.tpr -extend timetoextendby -o next.tpr
104 gmx mdrun -s next.tpr -cpi state.cpt
106 The time can also be extended using the ``-until`` and ``-nsteps``
107 options. Note that the original :ref:`mdp` file may have generated
108 velocities, but that is a one-time operation within :ref:`gmx grompp`
109 that is never performed again by any other tool.
111 Changing mdp options for a restart
112 ----------------------------------
114 If you wish to make changes to your simulations settings other than
115 length, then you should do so in the :ref:`mdp` file or topology, and
116 then call
118 ::
120 gmx grompp -f possibly-changed.mdp -p possibly-changed.top -c state.cpt -o new.tpr
121 gmx mdrun -s new.tpr -cpi state.cpt
123 to instruct :ref:`gmx grompp` to copy the full-precision coordinates
124 in the checkpoint file into the new :ref:`tpr` file. You should
125 consider your choices for :mdp:`tinit`, :mdp:`init-step`,
126 :mdp:`nsteps` and :mdp:`simulation-part`. You should generally not
127 regenerate velocities with :mdp:`gen-vel`, and generally select
128 :mdp:`continuation` so that constraints are not re-applied before
129 the first integration step.
131 Restarts without checkpoint files
132 ---------------------------------
134 It used to be possible to continue simulations without the checkpoint
135 files. As this approach could be unreliable or lead to
136 unphysical results, only restarts from checkpoints are permitted now.
138 Are continuations exact?
139 ------------------------
141 If you had a computer with unlimited precision, or if you integrated
142 the time-discretized equations of motion by hand, exact continuation
143 would lead to identical results. But since practical computers have
144 limited precision and MD is chaotic, trajectories will diverge very
145 rapidly even if one bit is different. Such trajectories will all be
146 equally valid, but eventually very different. Continuation using a
147 checkpoint file, using the same code compiled with the same compiler
148 and running on the same computer architecture using the same number of
149 processors without GPUs (see next section) would lead to binary
150 identical results. However,
151 by default the actual work load will be balanced across the hardware
152 according to the observed execution times. Such trajectories are
153 in principle not reproducible, and in particular a run that took
154 place in more than one part will not be identical with an equivalent
155 run in one part - but neither of them is better in any sense.
157 Reproducibility
158 ---------------
160 The following factors affect the reproducibility of a simulation, and thus its output:
162 * Precision (mixed / double) with double giving "better" reproducibility.
163 * Number of cores, due to different order in which forces are
164 accumulated. For instance (a+b)+c is not necessarily binary
165 identical to a+(b+c) in floating-point arithmetic.
166 * Type of processors. Even within the same processor family there can be slight differences.
167 * Optimization level when compiling.
168 * Optimizations at run time: e.g. the FFTW library that is typically
169 used for fast Fourier transforms determines at startup which version
170 of their algorithms is fastest, and uses that for the remainder of
171 the calculations. Since the speed estimate is not deterministic, the
172 results may vary from run to run.
173 * Random numbers used for instance as a seed for generating velocities
174 (in |Gromacs| at the preprocessing stage).
175 * Uninitialized variables in the code (but there shouldn't be any)
176 * Dynamic linking to different versions of shared libraries (e.g. for FFTs)
177 * Dynamic load balancing, since particles are redistributed to
178 processors based on elapsed wallclock time, which will lead to
179 (a+b)+c != a+(b+c) issues as above
180 * Number of PME-only ranks (for parallel PME simulations)
181 * MPI reductions typically do not guarantee the order of the
182 operations, and so the absence of associativity for floating-point
183 arithmetic means the result of a reduction depends on the order
184 actually chosen
185 * On GPUs, the reduction of e.g. non-bonded forces has a non-deterministic
186 summation order, so any fast implementation is non-reprodudible by
187 design.
189 The important question is whether it is a problem if simulations are
190 not completely reproducible. The answer is yes and no. Reproducibility
191 is a cornerstone of science in general, and hence it is important.
192 The `Central Limit Theorem <https://en.wikipedia.org/wiki/Central_limit_theorem>`_
193 tells us that in the case of infinitely long
194 simulations, all observables converge to their equilibrium
195 values. Molecular simulations in |Gromacs| adhere to this theorem, and
196 hence, for instance, the energy of your system will converge to a
197 finite value, the diffusion constant of your water molecules will
198 converge to a finite value, and so on. That means all the important
199 observables, which are the values you would like to get out of your
200 simulation, are reproducible. Each individual trajectory is not
201 reproducible, however.
203 However, there are a few cases where it would be useful if
204 trajectories were reproducible, too. These include developers doing
205 debugging, and searching for a rare event in a trajectory when, if
206 it occurs, you want to have manually saved your checkpoint file so
207 you can restart the simulation under different conditions, e.g.
208 writing output much more frequently.
210 In order to obtain this reproducible trajectory, it is important
211 to look over the list above and eliminate the factors that could
212 affect it. Further, using
214 ::
216 gmx mdrun -reprod
218 will eliminate all sources of non-reproducibility that it can,
219 i.e. same executable + same hardware + same shared libraries + same
220 run input file + same command line parameters will lead to
221 reproducible results.