LAMMPS Fenix process-recovery teaching workflow
Self-contained corresponding source bundles:Fenix develop fe22a98(2.0API) andLAMMPS30Sep2026 with explicit3-file lifecycle compatibility adapter;GPL/BSD notices retained.
No proprietary models,licensefiles,privateuserdata orruntimebinaries included.Build8coresmax into thisexample/runtime using existingreadonly OpenMPI5.0.10 andGNU13.3/C++20.
Existing sharedLAMMPSmodules/installations remainunchanged;this is anindependentexample runtime,not a claim allcluster binariesincludeFenix.
SourceC++adapter:excludeheldsparesfromnormalmodeuniverseconstruction;avoidfinalbarrieronstaleoriginalcomm;unwindstaleInputframeafterrecovery;knownsuccessfulspare/completed-recovery exits0.Othererrorsremainnonzero.
Unmodifiednative default-spareinitializationandpostrecoveryfinalization failures retainedprivately andnotqualified.
Own32atomFCC model,density.5,mass1,LJunits,epsilon1sigma1,cut1.8shifted,box4^3,initialT.1,COM0,fixedDATAIDs/velocities independentofMPIdecomposition.
NVE50stepsdt.001;savecheckpoint10;for recoverycases interrupt onlyoneverifiedownedrank1using uniqueper-launch token+UID+cwd+executable checks.
Fiveactualprotocols:plain4active;1spare3activecontrol;1spare3->3replacement;0spare4->3shrink;universalpartition3+1 withreplacement3->3.
Fresh MPIlauncher oncepercase;recoveryusesread_restart checkpoint10 andFenixrestart_label,recoveredmarker1,postrecoverydump beginsatstep10.
AllstrictnativeMPIreturncodes0;all6frames x32coordinates/velocities,energies,pressure/temp,independentpairforces checked.Originalgatesnotrelaxed.
Source275checks;energy drift4.6334e-7LJ;maxcoordinatevsplain1.774e-16;maxallframeforceformulaerror7.772e-15.
NoDFT/GPU orproductionnodefailure;theexampleterminatesonlytheownsmallMPIjob'sregisteredrank.This demonstratesprocess-level checkpointrecovery,not filesystem/powerfailure protection.
Viewfinishedreference:bash show-reference.sh;drawreference:bash plot-results.sh.
Recalculateinsideallocatedresources:sbatch job.slurm(buildsprivate runtime thenrunsfivecases);or bash prepare-runtime.sh thenbash run.sh.
Existingruntimeverifiedand reused;existingworkdirectoryrefused;noautomaticdeletion ofuserfiles.
MPIusesCPU ob1/self,tcp transport andULFM.Compatiblecompiler/OpenMPIrequired;do notmixMPIlibraries.
