NCP-AIO Workload Management Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Enable GPU time-slicing or MIG so multiple small jobs can share a device concurrently

  Why this is correct

  Time-slicing lets several containers share one physical GPU through rapid context switching, and MIG provides isolated slices; both allow small, bursty jobs to occupy a device concurrently instead of idling it during data-loading phases. This raises effective utilization without modifying model code, directly addressing the observed low utilization with queued work.

• ✗

  Move data loading to CPU-only nodes to eliminate the idle phases

  Why it's wrong here

  Data loading already occurs on CPU; the GPU idles while waiting for batches regardless of which node preprocesses them. Separating loaders onto CPU-only nodes adds network transfer overhead and does not let another job use the GPU during idle windows. It fails to address the core issue of exclusive device holding.

• ✗

  Add more GPU nodes to the cluster so queued jobs start sooner

  Why it's wrong here

  Adding nodes increases capacity but each job still monopolizes a full GPU while idling during data loading, so per-device utilization remains low and cost rises. The bottleneck is exclusive, inefficient device occupancy rather than insufficient hardware, so scaling out does not fix the underlying utilization problem.

• ✗

  Increase the batch size of each job so compute bursts last longer

  Why it's wrong here

  Changing batch size is a model-code or configuration change and often increases memory pressure, risking out-of-memory failures. It also does not address the fundamental issue that devices are held exclusively by jobs that idle during data loading. The requirement was to avoid model-code changes, so this action is inappropriate as a first step.