3.6 Analysis, restarts and HPC
The inputs on this page are educational templates that have not been executed for this course. Use an authorized installation, verify version-specific options and establish your own convergence.
3.6.1 An analysis contract
Before parsing document headers, atom order, units and periodicity. XYZ often lacks full cell/force metadata. Periodic bonds/diffusion need minimum images or unwrapping; subtracting wrapped positions can create jumps. Our nonperiodic water avoids that issue. Analyze geometry alongside energy/temperature, and do not count frames as independent samples: equilibration and correlation-aware errors are needed.
3.6.2 Electronic restart differs from trajectory continuation
water_sp-RESTART.wfn is an electronic guess; water_nve-1.restart can preserve positions, velocities, cell and motion state. The former alone does not continue the same MD trajectory. Inspect actual generated filenames and successful step, not merely a familiar name. EPCC restart practical.
# Inside DFT, alongside SCF:
WFN_RESTART_FILE_NAME ./water_sp-RESTART.wfn
# Inside the existing SCF block:
SCF_GUESS RESTART
Verify the intended wavefunction was loaded. Atom order, basis, spin, count, executable format or k representation may invalidate compatibility. An atomic guess can be a safer diagnostic after a model change. Restart reduces work but never licenses loose convergence.
For continuation copy complete restart and needed wavefunctions into a new segment. Carefully adapt a generated text restart's remaining steps and project/output names, or use EXT_RESTART in a new input:
# Top-level continuation block, after auditing saved state:
&EXT_RESTART
RESTART_FILE_NAME ./water_nve-1.restart
RESTART_DEFAULT .TRUE.
&END EXT_RESTART
Defaults restore many quantities including counters; audit inheritance before launching. External restart.
3.6.3 Diagnose the first cause
Missing data: actual scheduler directory, spelling, permissions and paths. Unknown basis/potential: exact entry in the actual read file, valence and XC; renaming cannot create an entry. SCF instability: geometry/contacts/charge/spin first, then fresh guess, grid and one solver change. Metals may need diagonalization/smearing/sampling rather than more OT. Keep failure detection on. MD drift/explosion: locate first bad step and previous frame, test timestep/force precision, units and accidental velocity reinitialization. Thermostats can conceal heating, not validate integration.
3.6.4 Benchmark HPC deliberately
MPI/OpenMP layout depends on hardware/system size. Use a representative larger model, keeping numerical settings fixed and recording time per converged step and total core-hours. Ranks×threads must fit allocation; separate scheduler memory/time errors from CP2K. Three-atom water cannot establish a production scaling optimum. EPCC performance exercise; ARCHER2 MPI/OpenMP workshop.
Use unique projects on suitable filesystems, limit unnecessary output at scale and copy irreplaceable inputs/logs/checkpoints to durable storage. Save scheduler scripts and executable versions. Public scripts must not contain passwords/private tokens. ARCHER2 CP2K guide is site-specific, not universal submission syntax.
3.6.5 Final assessment
Submit permitted inputs, single-point/convergence record, optimized coordinates, short MD segments and analysis script. Explain system, charge/spin/boundaries; executable/data files; grid/SCF/basis evidence; all geometry criteria; timestep diagnostic; restored state at segment boundary; and unsupported conclusions. Until run, label records planned and leave numerical fields blank. Further study: official manual and ARCHER2 university/HPC course.