Checkpointing and restart
DES3D writes checkpoint files at regular intervals so that a simulation can be resumed after interruption. With unchanged controls, the restart tests compare saved physical state with uninterrupted runs. Use one OpenMP thread for byte-exact PT continuation; multi-thread ADR reductions can change the last bits. The following state is persisted in the checkpoint:
dtanddt_PT(time step)max_global_vel_magreference_frame_timeinfo_display_next_stepdhacc(surface marker correction)strain_rate,viscosity,volume_oldplastic strain-rate(delta_plstrain), so a restarted run's diagnostic output matches a continuous one
Copy the original config and edit its existing [sim] section to select
a checkpoint explicitly, giving the restart a distinct output name:
[sim]
modelname = result-r
is_restarting = yes
restarting_from_modelname = result
restarting_from_frame = 10
Replace result and 10 with your original model prefix and an output
frame that has both save and checkpoint files, then run the executable
with that edited config. Checkpoint frame numbers are output-frame indices,
not necessarily physical step numbers. Retain the original physical
controls and set the stopping step/time to permit continuation. DES3D
requires both restart-selection parameters; rerunning an unchanged fresh
config does not automatically find a checkpoint.
PT restores saved velocity, next-step dt and velocity scale, and does not repeat completed initial equilibration. Its optional running scale requires the stored history; old checkpoints without it cannot silently start a new history. Hydraulic checkpoints also require the raw signed pressure increment and the saved ordering that determines whether it has already been consumed. Dry checkpoints do not require that hydraulic record. See PT restart and limits.
The checkpoint and frame file format was bumped from revision 3 to 4.
A revision (or dimensionality) mismatch between the file and the running
binary is now a hard error with a diagnostic naming the expected and
actual values, previously it was accepted silently and could be misread.
A checkpoint with the earlier revision cannot be restarted by the newer
solver. Preserve the original executable, configuration and data to continue
it. This restart restriction does not imply that every older saved field
is unreadable by Python: Dynearthsol.py reads frame metadata and available
fields separately, subject to the schema that reader supports.
Checkpoints now also carry nremesh, plstrain_remesh, and, in MMG builds,
surface dh and refined_elem_size_n (plus, with remeshing_option = 13,
side_prof_*). A checkpoint written before this change does not have these
records, and restarting from it is a restart-mismatch error (21/22).
Continue such checkpoints with their matching executable and preserve the
original data; the new state is not inferred automatically. See
Conservative remeshing
for what changed in MMG remeshing itself.
The optional finite-density hydraulic model also saves the physical inventories, numerical split buffer, EOS temperature origin and pending thermal/pressure increments. Accepted shear heat is saved as energy, so a changed timestep must not reinterpret an old power as a new energy. The current shear-heating checkpoint stores its version, pending extensive energy and accepted power together; missing members or an unsupported version are rejected. A power-only intermediate checkpoint is not converted. These model-specific guards apply in addition to the file-format revision. See Poromechanics.
Frame files (HDF5 and plain binary) now also embed the full .info row
(steps, time, dt, walltime, node/element/segment counts). If a run's
.info file is lost, utils/recreate_info.py rebuilds it from the frame
files alone; restart() also falls back to this embedded metadata
automatically when .info is missing.