Remote running, MPI and execution profiles
The same SPR-KKR calculations can run locally, over SSH, or through a job scheduler. Execution profiles select the machine, executable and execution backend without changing the scientific task.
KKRunner documentation describes the execution library. Its source is on GitHub.
Choosing an execution profile
Inspect the current settings and named profiles:
ase2sprkkr config --show
ase2sprkkr config --list-profiles
Without an explicitly selected named profile, ASE’s [sprkkr]
configuration supplies the defaults. See Installation and upgrading for
configuring local SPR-KKR executables.
For example, create a named profile for SSH execution without installing KKRunner on the remote host (replace the paths and host):
ase2sprkkr config --set profiles.cluster.executor ssh
ase2sprkkr config --set profiles.cluster.host user@cluster
ase2sprkkr config --set profiles.cluster.backend direct
ase2sprkkr config --set profiles.cluster.local_root /home/user/work
ase2sprkkr config --set profiles.cluster.remote_root /scratch/user/work
Run with that profile:
ase2sprkkr run calculation.pot --profile cluster
For an SSH job managed by a remote KKRunner supervisor, set
profiles.cluster.backend to kkrunner instead. This is also
the default backend for SSH profiles that do not specify a backend.
To make the named profile the default, run
ase2sprkkr config --set default_profile cluster.
Select a profile from Python with
calculator.calculate(profile="cluster"), or construct one directly:
from ase2sprkkr import SPRKKR
from ase2sprkkr.sprkkr.profile import SPRKKRProfile
profile = SPRKKRProfile(
executor="ssh",
host="user@cluster",
local_root="/path/to/local/work",
remote_root="/path/to/remote/work",
backend="direct",
)
result = SPRKKR(potential="calculation.pot").calculate(profile=profile)
Check that SSH login works, the remote directory is writable and the
remote host can run the required SPR-KKR and MPI executables. Set the
profile’s executable_dir if they are not on the remote PATH.
Execution backends and remote requirements
Direct (
backend="direct"): runs SPR-KKR over SSH using Python on the target, but does not require KKRunner there. It has no independent supervisor retaining a definitive exit status.KKRunner supervisor (
backend="kkrunner"): runs SPR-KKR under a supervisor on the execution host. This requires a compatible KKRunner installation in the target Python environment and retains completion information for reconnection and recovery.PSI/J (
backend="psij"): submits through a job scheduler, such as Slurm. Local scheduler access needsase2sprkkr[psij]on the client; scheduler access over SSH requires a compatiblekkrunner[psij]installation on the remote host. Setpsij_backend="slurm"for Slurm.
These are backend selections, independent of whether the calculation is recorded in a project. See the KKRunner backend guide for further details.
Persistence, reattachment and recovery
These describe different properties of a run:
Persistence: saving a run ID, its status, file paths and other metadata in a run store. Any backend can be used with a persistent project. Untracked execution is also possible with any backend.
Reattachment: reconnecting to a known run after the original client session ends. With a direct backend, a saved process identity and output paths allow limited status checks and resuming observation of output. Once the original connection is lost, however, the final exit code may be unavailable, so completion cannot necessarily be verified as successful. The
kkrunnersupervisor retains completion information and supports fuller reattachment.Submission recovery: finding a job when the client disconnected before its backend ID was saved. Direct execution cannot do this. The supervisor or a scheduler may support it, subject to available receipts or scheduler records.
A lost connection does not, by itself, prove that a running process has terminated. Equally, a persistent record alone does not guarantee reliable reattachment or successful postprocessing. See Persistent project calculations for enabling project tracking and loading history.
MPI and executable selection
MPI defaults to auto. Request four ranks with:
from ase2sprkkr import SPRKKR
result = SPRKKR(potential="calculation.pot", mpi=4).calculate()
A required MPI calculation must not silently fall back to serial. If only serial executables are installed and automatic selection cannot use them, disable MPI:
ase2sprkkr config --set profiles.default.mpi false
For Slurm, select backend="psij" and
psij_backend="slurm". With executor="psij", PSI/J
connects to a local scheduler; with executor="ssh", the
backend runs on a scheduler-accessible host via SSH.
Monitoring recorded runs
Use ase2sprkkr remote status, follow and cancel
with saved run IDs. The --remote-id alternative accepts an
explicit backend handle instead. See Command-line reference for syntax.