McAfee-Secured Website

Exam Code: NCP-AII

Exam Name: NCP-AI Infrastructure

Certification Provider: NVIDIA

Questions & Answers

100% NVIDIA NCP-AII Guaranteed Success With Testking Training Materials

Get Real NCP-AII Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

Questions & Answers

170 Questions & Answers with Testing Engine

"NCP-AI Infrastructure Exam", also known as NCP-AII exam, is a NVIDIA certification exam.

Pass your tests with the always up-to-date NCP-AII Exam Engine. Your NCP-AII training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

NVIDIA NCP-AII Practice Test 27

Question 260

During a GPU datacenter maintenance window, an engineer discovers: Slurm reports some nodes as allocated indefinitely after a failed distributed job because job cleanup leaves Enroot container artifacts and stale GPU processes. Which repeatable operational safeguard best prevents recurrence of this failure?

  1. Review job launch templates and interface naming after rack hardware and operating system refreshes.
  2. Require access-policy review and negative scheduling tests before adding new GPU resource pools.
  3. Validate least-privilege NGC account access and secret custody before admitting shared management hosts.
  4. Exercise failed-job cleanup and account isolation in the cluster integration acceptance suite.

Explanation

Correct Answer

D:

This preventive practice addresses the underlying failure (Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.) and supports continued assurance through: A deliberately interrupted test job frees its GPU assignments and leaves the host ready for a new tenant after verified cleanup.

Incorrect Answers

A:

This prevention control targets another operational risk: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. The currently reported mechanism remains: Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.

B:

This prevention control targets another operational risk: A new node is added to the wrong Slurm partition and gains access to restricted customer workloads despite being intended for untrusted experimentation. The currently reported mechanism remains: Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.

C:

This prevention control targets another operational risk: An NGC CLI client is installed on a shared administrative workstation but lacks authenticated registry configuration, so container downloads fail with permissions errors. The currently reported mechanism remains: Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.

Question 261

A GPU cluster commissioning engineer investigates this situation: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. Which corrective deployment or operational action best addresses the stated problem?

  1. Verify supported Enroot and Pyxis deployment and Slurm plugin configuration across controller and compute nodes.
  2. Apply supported job-epilog cleanup and node health checks that terminate only abandoned workloads and release scheduler resources safely.
  3. Align host DOCA, BlueField software, and firmware to a vendor-supported compatibility matrix and upgrade through a controlled sequence.
  4. Normalize node network interface configuration or select supported per-node fabric parameters through managed job profiles.

Explanation

Correct Answer

D:

For the stated NVIDIA AI infrastructure condition (A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack.), this action addresses the concrete failure: Normalize node network interface configuration or select supported per-node fabric parameters through managed job profiles. The underlying mechanism is: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent. Verification would require: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.

Incorrect Answers

A:

This option could be appropriate for a different NVIDIA infrastructure problem (3.3: Install clusters, categories, interfaces, Slurm, Enroot, and Pyxis): A training cluster stages a container with Enroot, but Slurm launches it without the intended Pyxis integration and the job cannot access accelerators. The present issue is more specifically caused by: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

B:

This option could be appropriate for a different NVIDIA infrastructure problem (3.3: Install clusters, categories, interfaces, Slurm, Enroot, and Pyxis): Slurm reports some nodes as allocated indefinitely after a failed distributed job because job cleanup leaves Enroot container artifacts and stale GPU processes. The present issue is more specifically caused by: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

C:

This option could be appropriate for a different NVIDIA infrastructure problem (3.4: Install, update, and remove NVIDIA GPU and DOCA drivers): A BlueField DPU application fails after the host updates DOCA packages, while the DPU firmware and embedded software versions remain from an earlier release. The present issue is more specifically caused by: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

Question 262

A GPU cluster commissioning engineer investigates this situation: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. Which measured result most directly proves that this specific problem is resolved?

  1. Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.
  2. The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.
  3. A sample GPU job requests the expected accelerator count and Slurm allocates only the authorized capacity on a healthy node.
  4. A small container runs the approved `nvidia-smi` or CUDA probe and reports the allocated device UUIDs correctly.

Explanation

Correct Answer

B:

This is the measurable acceptance result for the described case: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target. It validates the corrective operation (Normalize node network interface configuration or select supported per-node fabric parameters through managed job profiles.) rather than a related but different system check.

Incorrect Answers

A:

This observation would help verify a different incident: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. It does not directly establish the required result for the current case: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.

C:

This observation would help verify a different incident: An AI job runs on one compute node, but Slurm cannot distribute it because GPU resources are absent from the scheduler's generic resource definitions. It does not directly establish the required result for the current case: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.

D:

This observation would help verify a different incident: A Docker AI image launches on a GPU host but reports no CUDA devices because the NVIDIA container runtime integration was never installed. It does not directly establish the required result for the current case: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.

Question 263

A GPU cluster commissioning engineer investigates this situation: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. Which root cause or infrastructure principle best explains the observed failure?

  1. Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.
  2. GPUs visible to the OS are not automatically discoverable by the cluster scheduler without correctly mapped resources.
  3. An installed host driver does not automatically expose accelerators inside isolated containers.
  4. A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

Explanation

Correct Answer

D:

This technical mechanism explains the reported condition: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent. The observed situation is: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. A requirement-aligned correction is: Normalize node network interface configuration or select supported per-node fabric parameters through managed job profiles.

Incorrect Answers

A:

This underlying cause is associated with another GPU platform incident: Slurm reports some nodes as allocated indefinitely after a failed distributed job because job cleanup leaves Enroot container artifacts and stale GPU processes. The observed case instead concerns: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

B:

This underlying cause is associated with another GPU platform incident: An AI job runs on one compute node, but Slurm cannot distribute it because GPU resources are absent from the scheduler's generic resource definitions. The observed case instead concerns: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

C:

This underlying cause is associated with another GPU platform incident: A Docker AI image launches on a GPU host but reports no CUDA devices because the NVIDIA container runtime integration was never installed. The observed case instead concerns: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

Question 264

A GPU cluster commissioning engineer investigates this situation: A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. Which repeatable operational safeguard best prevents recurrence of this failure?

  1. Include scheduler-to-container launch smoke tests in every image and Slurm plugin upgrade gate.
  2. Require access-policy review and negative scheduling tests before adding new GPU resource pools.
  3. Review job launch templates and interface naming after rack hardware and operating system refreshes.
  4. Enforce job draining, module state checks, and postchange CUDA health gates before driver maintenance.

Explanation

Correct Answer

C:

This preventive practice addresses the underlying failure (A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.) and supports continued assurance through: The distributed test uses the intended high-speed interfaces on all ranks and passes an inter-node collective bandwidth target.

Incorrect Answers

A:

This prevention control targets another operational risk: A training cluster stages a container with Enroot, but Slurm launches it without the intended Pyxis integration and the job cannot access accelerators. The currently reported mechanism remains: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

B:

This prevention control targets another operational risk: A new node is added to the wrong Slurm partition and gains access to restricted customer workloads despite being intended for untrusted experimentation. The currently reported mechanism remains: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

D:

This prevention control targets another operational risk: An operator removes the NVIDIA driver while GPU containers are still running, leaving orphaned jobs and an inconsistent module state across the node. The currently reported mechanism remains: A scheduler can place jobs correctly yet fail at communication if its node-level fabric assumptions are inconsistent.

Question 265

During NVIDIA AI factory acceptance, a technician observes: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. Which corrective deployment or operational action best addresses the stated problem?

  1. Normalize node network interface configuration or select supported per-node fabric parameters through managed job profiles.
  2. Place scheduler management on a supported resilient control-plane architecture and test recovery independently of compute fabric failures.
  3. Verify supported Enroot and Pyxis deployment and Slurm plugin configuration across controller and compute nodes.
  4. Align host DOCA, BlueField software, and firmware to a vendor-supported compatibility matrix and upgrade through a controlled sequence.

Explanation

Correct Answer

B:

For the stated NVIDIA AI infrastructure condition (An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack.), this action addresses the concrete failure: Place scheduler management on a supported resilient control-plane architecture and test recovery independently of compute fabric failures. The underlying mechanism is: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks. Verification would require: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.

Incorrect Answers

A:

This option could be appropriate for a different NVIDIA infrastructure problem (3.3: Install clusters, categories, interfaces, Slurm, Enroot, and Pyxis): A multi-node NCCL job spans nodes with different fabric interface names, while Slurm launch scripts hard-code the interface present only on the first rack. The present issue is more specifically caused by: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

C:

This option could be appropriate for a different NVIDIA infrastructure problem (3.3: Install clusters, categories, interfaces, Slurm, Enroot, and Pyxis): A training cluster stages a container with Enroot, but Slurm launches it without the intended Pyxis integration and the job cannot access accelerators. The present issue is more specifically caused by: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

D:

This option could be appropriate for a different NVIDIA infrastructure problem (3.4: Install, update, and remove NVIDIA GPU and DOCA drivers): A BlueField DPU application fails after the host updates DOCA packages, while the DPU firmware and embedded software versions remain from an earlier release. The present issue is more specifically caused by: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

Question 266

During NVIDIA AI factory acceptance, a technician observes: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. Which measured result most directly proves that this specific problem is resolved?

  1. Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.
  2. A scheduled canary launches the approved container and accesses its allocated GPUs with successful startup logs and correct permissions.
  3. A denied test account cannot schedule into the restricted partition while an approved account receives its expected GPU allocation.
  4. The postupgrade test provisions a canary GPU node and completes a scheduled workload without losing existing management or node state.

Explanation

Correct Answer

A:

This is the measurable acceptance result for the described case: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational. It validates the corrective operation (Place scheduler management on a supported resilient control-plane architecture and test recovery independently of compute fabric failures.) rather than a related but different system check.

Incorrect Answers

B:

This observation would help verify a different incident: A training cluster stages a container with Enroot, but Slurm launches it without the intended Pyxis integration and the job cannot access accelerators. It does not directly establish the required result for the current case: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.

C:

This observation would help verify a different incident: A new node is added to the wrong Slurm partition and gains access to restricted customer workloads despite being intended for untrusted experimentation. It does not directly establish the required result for the current case: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.

D:

This observation would help verify a different incident: An operations team upgrades BCM directly in production without testing whether existing Slurm and provisioning integrations are compatible with the new management API. It does not directly establish the required result for the current case: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.

Question 267

During NVIDIA AI factory acceptance, a technician observes: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. Which root cause or infrastructure principle best explains the observed failure?

  1. Stale workloads and container state can make a healthy node appear occupied or leak resource ownership between jobs.
  2. Cluster resource categories affect authorization as well as performance, and incorrect assignment can cross tenant boundaries.
  3. GPU container portability has a dependency on host driver capabilities beyond the image's packaged libraries.
  4. Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

Explanation

Correct Answer

D:

This technical mechanism explains the reported condition: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks. The observed situation is: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. A requirement-aligned correction is: Place scheduler management on a supported resilient control-plane architecture and test recovery independently of compute fabric failures.

Incorrect Answers

A:

This underlying cause is associated with another GPU platform incident: Slurm reports some nodes as allocated indefinitely after a failed distributed job because job cleanup leaves Enroot container artifacts and stale GPU processes. The observed case instead concerns: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

B:

This underlying cause is associated with another GPU platform incident: A new node is added to the wrong Slurm partition and gains access to restricted customer workloads despite being intended for untrusted experimentation. The observed case instead concerns: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

C:

This underlying cause is associated with another GPU platform incident: A container sees its host GPU but fails when run on a different node whose driver is older than the image's required CUDA feature level. The observed case instead concerns: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

Question 268

During NVIDIA AI factory acceptance, a technician observes: An AI cluster experiences scheduler outages whenever a compute rack is disconnected because the Slurm controller itself resides on a nonredundant compute host in that rack. Which repeatable operational safeguard best prevents recurrence of this failure?

  1. Require access-policy review and negative scheduling tests before adding new GPU resource pools.
  2. Include scheduler-to-container launch smoke tests in every image and Slurm plugin upgrade gate.
  3. Document scheduler control-plane placement and failover tests separately from GPU resource high availability.
  4. Approve minimal GPU container permissions and prohibit privilege escalation as a default deployment shortcut.

Explanation

Correct Answer

C:

This preventive practice addresses the underlying failure (Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.) and supports continued assurance through: Isolation of the affected compute rack reduces available GPUs but leaves job administration and healthy rack scheduling operational.

Incorrect Answers

A:

This prevention control targets another operational risk: A new node is added to the wrong Slurm partition and gains access to restricted customer workloads despite being intended for untrusted experimentation. The currently reported mechanism remains: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

B:

This prevention control targets another operational risk: A training cluster stages a container with Enroot, but Slurm launches it without the intended Pyxis integration and the job cannot access accelerators. The currently reported mechanism remains: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

D:

This prevention control targets another operational risk: A GPU-based inference container is deployed with `--privileged` solely to make NVIDIA device files visible, exposing unnecessary host capabilities. The currently reported mechanism remains: Control services tied to ordinary compute failure domains create avoidable cluster-wide availability risks.

Question 269

An infrastructure administrator receives the following incident: A new NVIDIA driver is installed on every host without checking its supported CUDA runtime and kernel module compatibility, leading to failed CUDA initialization. Which corrective deployment or operational action best addresses the stated problem?

  1. Verify driver module signing and compatibility with the approved Secure Boot and kernel policy before reinstalling packages.
  2. Drain the scheduler node, stop GPU workloads, and use the supported driver removal or update workflow followed by validation.
  3. Choose a qualified driver/runtime combination and validate kernel module loading on a canary before fleet rollout.
  4. Design and verify BCM high availability with supported redundant management services and resilient state storage.

Explanation

Correct Answer

C:

For the stated NVIDIA AI infrastructure condition (A new NVIDIA driver is installed on every host without checking its supported CUDA runtime and kernel module compatibility, leading to failed CUDA initialization.), this action addresses the concrete failure: Choose a qualified driver/runtime combination and validate kernel module loading on a canary before fleet rollout. The underlying mechanism is: GPU driver releases have compatibility requirements that an otherwise healthy Linux boot does not guarantee. Verification would require: A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation.

Incorrect Answers

A:

This option could be appropriate for a different NVIDIA infrastructure problem (3.4: Install, update, and remove NVIDIA GPU and DOCA drivers): A security patch changes the host kernel and Secure Boot settings, and the NVIDIA GPU module stops loading despite the device remaining visible on PCIe. The present issue is more specifically caused by: GPU driver releases have compatibility requirements that an otherwise healthy Linux boot does not guarantee.

B:

This option could be appropriate for a different NVIDIA infrastructure problem (3.4: Install, update, and remove NVIDIA GPU and DOCA drivers): An operator removes the NVIDIA driver while GPU containers are still running, leaving orphaned jobs and an inconsistent module state across the node. The present issue is more specifically caused by: GPU driver releases have compatibility requirements that an otherwise healthy Linux boot does not guarantee.

D:

This option could be appropriate for a different NVIDIA infrastructure problem (3.1: Install Base Command Manager and configure and verify high availability): Base Command Manager is installed on a management node but the cluster database is stored only on a single local disk, and leadership claims the control plane is highly available. The present issue is more specifically caused by: GPU driver releases have compatibility requirements that an otherwise healthy Linux boot does not guarantee.

Question 270

An infrastructure administrator receives the following incident: A new NVIDIA driver is installed on every host without checking its supported CUDA runtime and kernel module compatibility, leading to failed CUDA initialization. Which measured result most directly proves that this specific problem is resolved?

  1. A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation.
  2. The package manifest identifies trusted origin, the driver loads cleanly, and a representative workload completes with recorded versions.
  3. A supported DOCA sample completes end-to-end and both host and DPU report the approved paired software baseline.
  4. A controlled loss of the primary management node preserves or restores administrative provisioning through the supported standby within the accepted recovery target.

Explanation

Correct Answer

A:

This is the measurable acceptance result for the described case: A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation. It validates the corrective operation (Choose a qualified driver/runtime combination and validate kernel module loading on a canary before fleet rollout.) rather than a related but different system check.

Incorrect Answers

B:

This observation would help verify a different incident: GPU health tests fail because the driver package on a server was installed from an untrusted repository, and management cannot reproduce its package dependencies. It does not directly establish the required result for the current case: A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation.

C:

This observation would help verify a different incident: A BlueField DPU application fails after the host updates DOCA packages, while the DPU firmware and embedded software versions remain from an earlier release. It does not directly establish the required result for the current case: A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation.

D:

This observation would help verify a different incident: Base Command Manager is installed on a management node but the cluster database is stored only on a single local disk, and leadership claims the control plane is highly available. It does not directly establish the required result for the current case: A known CUDA sample and `nvidia-smi` both run successfully on the approved kernel and the remaining fleet retains known-good operation.

NCP-AI Infrastructure (NCP-AII) Practice Tests