Cluster Commander
Schedule jobs without fragmenting the fleet.
How to play
You have an 8x8 GPU cluster. Jobs appear in the queue with varying shapes (1x1 inference, 2x2 fine-tuning, 4x4 pre-training). Click a valid empty space in the grid to schedule the next job. If the grid becomes too fragmented to fit a large job, it will block the entire queue.
The Systems Concept
Scheduling diverse jobs on a shared GPU cluster (like Slurm or Kubernetes) often leads to fleet fragmentation. When many small jobs scatter across the cluster, a large contiguous job (e.g., a large pre-training run requiring 16 interconnected GPUs) might be unable to start, causing cluster utilization to drop sharply. Modern schedulers use defragmentation and backfilling to mitigate this.
Part of MLSysBook Playground. Found a bug? Report an issue.