Output, restart and resume¶

What a run writes¶

fm.Output("run/", trajectory_every=20, thermo_every=20, restart_every=10_000) (the fm.md defaults) writes into run/:

File	Content
`thermo.<n>.csv`	thermo rows (step, time, temperature, energies, pressure for NPT), each chunk with its header
`trajectory.<n>.traj`	ASE trajectory frames
`restart.npz`	the complete simulation state at the last restart point
`simulation.json`	the inputs and planner options, for `resume`
`metadata.json`	plan, gate, report, provenance (versions, model hash, GPU)

The outputs are written in chunks, n = 0, 1, ..., with a chunk boundary at every restart point. Each chunk is complete on its own: ase.io.read("run/trajectory.0.traj", index=":") reads one. To read a whole run in order:

from fastmace.io.output import read_frames, read_thermo

frames = read_frames("run/")
thermo = read_thermo("run/")

Output is asynchronous: a snapshot is copied to the host and written while the next step runs, so writing does not slow the GPU.

Restart¶

A restart file holds the complete state: positions, momenta, cell, barostat and thermostat state, the RNG state, species after MC exchanges, the schedule position, the MC counters and the plan revision. Restart points are written every restart_every steps, at the end of each protocol piece and by save_restart. A run whose steps is not a multiple of restart_every has no restart at its last step; see Output files.

sim.save_restart("run/restart.npz")         # an extra one by hand
sim = fm.Simulation.resume("run/")          # continue
sim.run()                                   # the rest of the schedule

Resume checks the chunks up to the restart point, deletes later chunks (rows and frames written after the restart point) and continues with the next chunk. Nothing is rewritten. If the stored chunks do not match, it raises RestartMismatch before touching anything.

A resume rebuilds the recorded plan. If the software or GPU changed so that the plan would differ, it refuses unless you pass allow_plan_change=True. A resume never falls back.

Reproducibility¶

Hybrid runs (the default backend) use atomic sums and are not bitwise reproducible; compare them statistically. A resumed run agrees bit for bit with an uninterrupted one in plans made with fm.Workload(reproducible=True), which the planner refuses where it cannot guarantee it. Restart points rebuild the neighbour list, so for bitwise agreement set restart_every=None too.

Command line¶

fastmace run run.toml writes the same files. To quench frames of a finished run, see Relaxations and quenches.