The most expensive state a fleet can be in: the scheduler counts the device as taken, the bill counts it as running, and nothing is using it.
1. Find themOpen the Overview. Devices that are allocated and idle are flagged there, and the fleet line counts them. On the command line: siltide --tab overview --filter "util<5 procs>0".
2. Confirm it is not a gap in measurementPress 8 for History and scrub back with ,. A device that has been flat for the whole window is idle; one that dips between bursts is a training loop between steps.
3. Find who holds itPress enter on the device. The process table names the pid, the user, and the pod or Slurm job behind it.
4. Check the other halfOn the Nodes tab, the scheduler table says how many devices the cluster handed out on that node. Eight handed out and three idle is the number to take to whoever owns the workload.
5. DecideA notebook left open, a job that finished but did not exit, or a pod requesting more devices than it uses. siltide will not act on any of them: it tells you which one it is.