問題文
A cluster runs many single-accelerator jobs, and the operations team finds that a node with eight accelerators is occupied by one such job while seven devices sit idle. They want several jobs to share a node without oversubscribing anything. What changes this behavior?
選択肢
- Ask users to submit with the exclusive option removed from their scripts.
- Switch the selection plugin to the one that tracks consumable resources, which lets jobs be co-scheduled on a node while still accounting for every unit that is dedicated to each job.
- Raise the oversubscription setting on the partition, which is what allows more than one job to occupy a node and is the only supported way to reach that state when jobs ask for whole devices.
- Lower the default time limit so that jobs finish sooner and the idle devices are reached by the next job in the queue.