What it is
GPU clusters rarely fail for lack of GPUs alone. Free cards end up on the wrong nodes, split across NVLink islands or held by half-placed multi-pod jobs, while big jobs queue. This simulator runs one seeded workload through several scheduling policies and measures the damage: fragmentation, queue wait, utilization and preemption cost.
When to use it
Use it to see why a policy choice changes wait times, to size a checkpoint interval against preemption, or to explain gang scheduling and quota borrowing to a team. Compare policies on identical jobs, then replay one on a timeline and scrub through it.
Gotchas & limits
These are teaching models. kube-default and binpack placements were checked against the K8s Scheduler Simulator; gang and quota queues follow the ideas of Volcano and Kueue but are not their implementations, and the speed penalties for spreading a job are illustrative. Results depend on the generated workload, so compare policies on the same seed. A run that leaves jobs waiting at the end has no makespan.