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.

• ✗

  Add a tolerations entry for nvidia.com/gpu to each pod and increase the kubelet's --max-pods flag on GPU nodes.

  Why it's wrong here

  Tolerations only affect scheduling against taints and have no bearing on how many pods can share a GPU. Raising --max-pods increases the number of pods a node can run, but the GPU remains a single allocatable extended resource, so the second pod requesting nvidia.com/gpu would still be unschedulable. Neither change enables sharing of one device.

• ✗

  Set the environment variable NVIDIA_VISIBLE_DEVICES=all on each pod and let the runtime divide GPU time among containers.

  Why it's wrong here

  Setting NVIDIA_VISIBLE_DEVICES=all exposes every GPU to the container but does not create replicas or enforce time-slicing. The device plugin still advertises only one allocatable GPU per physical device, so the scheduler cannot place multiple pods on the same GPU. Time-slicing must be configured at the device plugin level, not through per-pod environment variables.

• ✓

  Create a ConfigMap containing the time-slicing configuration and reference it in the ClusterPolicy so the GPU Operator propagates the device plugin config across nodes.

  Why this is correct

  The GPU Operator supports time-slicing by reading a ConfigMap referenced in the ClusterPolicy's device plugin configuration. When applied, the operator propagates the config to all GPU nodes and restarts the device plugin, causing each physical GPU to advertise a multiplied replica count so multiple pods can share one device. This is the supported cluster-wide method.

• ✗

  Install the NVIDIA MIG Manager and set the nvidia.com/mig.config label to all-1g.5gb on each node to subdivide the GPUs.

  Why it's wrong here

  MIG partitions a GPU into separate hardware-isolated instances, not time-sliced shares. Requesting mig-1g.5gb profiles gives each pod a dedicated slice with fixed memory, which changes both isolation semantics and the resource name. MIG is a distinct mechanism from time-slicing and does not let multiple pods share the same GPU partition concurrently.