Running one simulation is easy. Running dozens of variants of it, and remembering three weeks later which output folder was the one with the finer grid and the lower viscosity, is not. MFC writes its outputs into the case directory, where the next run overwrites them. I wrote mfc-run in February 2026 so that every run keeps a record of what it was for and where its results went.
What it does
mfc-run wraps MFC’s ./mfc.sh run. Every run gets a short description, and when it finishes, everything it made is filed into its own dated folder:
run_info.json, recording the status, duration, MFC git commit and parameters- a snapshot of
case.pyas it was when the run started, plus the parameters as JSON - the simulation outputs (
D/,silo_hdf5/,restart_data/,p_all/), moved out of the working directory - copies of any analysis scripts and plots
Failed runs go into a separate failed/ folder instead of being mixed in with good ones, and mfc-run --list-runs <case> shows a case’s whole history. It finds the MFC checkout on its own by walking up from the case file. It’s a single Bash script that needs nothing beyond Python, rsync and git, so it drops onto an HPC cluster easily.
Batch jobs are where it gets interesting
On a workstation, a run is finished when the command returns. On a cluster it isn’t. The first version of batch support archived empty results, because sbatch returns immediately after submitting a job, long before the job runs. The fix was to poll the scheduler until the job actually finished before archiving anything.
A later audit found two subtler problems. A still-running batch job could be recorded as an empty success. And the scheduler settings the script was exporting did nothing at all, because MFC reads its account and partition from its own command-line flags, not from the environment. The reworked batch mode supports SLURM, PBS and LSF, only deletes local outputs after a clean success, and comes with a 220-line test suite that fakes each scheduler so it can be tested without a cluster. That rework is written and tested but not merged yet.
Upstream
The same idea later went into MFC itself as the --archive option for interactive runs (#1376), which I wrote and which was merged in April 2026.