10 free NVIDIA NCP-AIO practice questions with the correct answer and a full explanation for each, taken from the CertStash pack of 66 questions. Work through them, then open each answer to check your reasoning.
Get all 66 questions (US$39) · Download these 10 as a PDF
Question 1
A Slurm user needs to submit a batch job script for execution tomorrow.
Which command should be used to complete this task?
Show answer and explanation
Correct answer: A. sbatch -begin=tomorrow
sbatch is the Slurm command for submitting batch job scripts for deferred execution. The –begin parameter (or -begin) allows users to specify when a job should start, including tomorrow or any future time. submit is not a valid Slurm command, salloc allocates resources interactively rather than submitting batch jobs, and srun executes commands directly without batch submission.
Why the other options are wrong
- B. submit is not a valid Slurm command; Slurm uses sbatch for batch job submission.
- C. salloc is used for interactive resource allocation, not batch job submission with deferred execution.
- D. srun executes commands directly on allocated resources without submitting batch jobs to the queue.
Question 2
You are configuring networking for a new AI cluster in your data center. The cluster will handle large-scale distributed training jobs that require fast communication between servers.
What type of networking architecture can maximize performance for these AI workloads?
Show answer and explanation
Correct answer: D. Use InfiniBand networking to provide low-latency, high-throughput communication between servers in the cluster.
high-throughput communication between servers in the cluster. InfiniBand networking provides the low-latency, high-throughput communication required for distributed AI training workloads where nodes need to exchange large amounts of data rapidly. InfiniBand's superior performance characteristics make it the optimal choice for AI clusters compared to standard Ethernet. While leaf-spine topology (option A) is a good architectural pattern, it applies to Ethernet networks. Out-of-band management networks (option B) are separate from compute networks and less critical than compute performance. Standard Ethernet with multiple connections (option C) cannot match InfiniBand's latency and throughput for this use case.
Why the other options are wrong
- A. Leaf-spine topology with standard Ethernet is scalable but does not provide the low- latency, high-throughput performance needed for AI distributed training compared to InfiniBand.
- B. Out-of-band management networks are administrative, not compute networks; prioritizing them over compute networks would harm AI workload performance.
- C. Standard Ethernet with increased bandwidth cannot achieve the low-latency performance characteristics that InfiniBand provides for distributed AI training.
Question 3
A system administrator needs to optimize the delivery of their AI applications to the edge.
What NVIDIA platform should be used?
Show answer and explanation
Correct answer: C. Fleet Command
Fleet Command is NVIDIA's platform specifically designed for deploying and managing AI applications at the edge. It provides orchestration, lifecycle management, and optimization of edge AI applications across distributed edge nodes. Base Command Platform and Manager are designed for data center-scale AI cluster management and job scheduling. NetQ is NVIDIA's network telemetry and management tool focused on network infrastructure monitoring.
Why the other options are wrong
- A. Base Command Platform is for managing large-scale data center AI clusters, not edge deployment optimization.
- B. Base Command Manager handles cluster resource management and job scheduling for data centers, not edge AI application delivery.
- D. NetQ is a network visibility and telemetry tool for infrastructure monitoring, not an AI application delivery platform.
Question 4
A Slurm user is experiencing a frequent issue where a Slurm job is getting stuck in the “PENDING” state and unable to progress to the “RUNNING” state.
Which Slurm command can help the user identify the reason for the job’s pending status?
Show answer and explanation
Correct answer: B. scontrol show job <jobid></jobid>
The scontrol show job command displays detailed information about a specific job, including its current state, resource requirements, scheduling constraints, and reasons why it may be pending. This command provides the diagnostic information needed to determine why a job is stuck in PENDING state. sinfo -R shows node reasons but not job- specific details, sacct is for historical job accounting after completion, and squeue shows job status overview without detailed pending reasons.
Why the other options are wrong
- A. sinfo -R displays node reservation and down reasons, not job-specific pending reasons.
- C. sacct shows historical accounting information for completed jobs, not current pending job details.
- D. squeue displays a brief queue status overview without detailed information about why specific jobs are pending.
Question 5
You are a Solutions Architect designing a data center infrastructure for a cloud-based AI application that requires high-performance networking, storage, and security. You need to choose a software framework to program the NVIDIA BlueField DPUs that will be used in the infrastructure. The framework must support the development of custom applications and services, as well as enable tailored solutions for specific workloads.
Additionally, the framework should allow for the integration of storage services such as NVMe over Fabrics (NVMe-oF) and elastic block storage.
Which framework should you choose?
Show answer and explanation
Correct answer: D. NVIDIA DOCA
NVIDIA DOCA (Data Center AI and Accelerated Computing) is the framework specifically designed for programming BlueField DPUs. DOCA enables development of custom applications, data center services, and supports integration with storage services including NVMe-oF and block storage, making it ideal for building tailored solutions for specific workloads on DPUs. TensorRT is for model inference optimization, CUDA is for GPU computing, and NSight is a profiling and debugging tool.
Why the other options are wrong
- A. TensorRT is for optimizing deep learning model inference on GPUs and GPUs, not for BlueField DPU application development.
- B. CUDA is for GPU parallel computing, not for programming BlueField DPUs which require DPU-specific frameworks.
- C. NSight is a profiling, debugging, and analysis tool, not a development framework for writing DPU applications.
Question 6
You are managing a Slurm cluster with multiple GPU nodes, each equipped with different types of GPUs. Some jobs are being allocated GPUs that should be reserved for other purposes, such as display rendering.
How would you ensure that only the intended GPUs are allocated to jobs?
Show answer and explanation
Correct answer: A. Verify that the GPUs are correctly listed in both gres.conf and slurm.conf, and ensure that unconfigured GPUs are excluded.
gres.conf and slurm.conf, and ensure that unconfigured GPUs are excluded. Proper GPU allocation control in Slurm requires that GPUs be correctly configured in both gres.conf (Generic Resource configuration) and slurm.conf (main Slurm configuration). Additionally, any GPUs not intended for job allocation must be explicitly excluded from the configuration so they remain unavailable to the scheduler. This ensures only intended GPUs are allocated. Manual assignment with nvidia-smi (option B) is not scalable, driver reinstallation (option C) does not control allocation, and requesting more GPUs (option D) wastes resources and contradicts the intent.
Why the other options are wrong
- B. Manually assigning GPUs with nvidia-smi is not a scalable cluster management solution and defeats the purpose of resource scheduling.
- C. Reinstalling NVIDIA drivers addresses driver issues but does not control which GPUs Slurm allocates to jobs.
- D. Requesting more GPUs than needed increases resource waste and contention, not preventing unintended GPU allocation.
Question 7
A data scientist is training a deep learning model and notices slower than expected training times. The data scientist alerts a system administrator to inspect the issue. The system administrator suspects the disk IO is the issue.
What command should be used?
Show answer and explanation
Correct answer: B. iostat
iostat is the standard Linux command for monitoring and analyzing disk I/O performance statistics including read/write rates, utilization, and wait times. It directly provides the metrics needed to diagnose disk I/O bottlenecks affecting training performance. tcpdump analyzes network traffic, nvidia-smi monitors GPU status, and htop shows CPU and memory usage but not detailed I/O statistics.
Why the other options are wrong
- A. tcpdump captures network packets for network analysis, not disk I/O performance monitoring.
- C. nvidia-smi displays GPU status and memory usage, not disk I/O performance metrics.
- D. htop shows CPU and memory utilization but does not provide detailed disk I/O statistics needed to diagnose storage bottlenecks.
Question 8
You have noticed that users can access all GPUs on a node even when they request only one GPU in their job script using –gres=gpu:1. This is causing resource contention and inefficient GPU usage.
What configuration change would you make to restrict users’ access to only their allocated GPUs?
Show answer and explanation
Correct answer: B. Enable cgroup enforcement in cgroup.conf by setting ConstrainDevices=yes.
ConstrainDevices=yes. Enabling cgroup enforcement in Slurm's cgroup.conf through the ConstrainDevices=yes setting restricts processes within a job to access only the resources (including GPUs) explicitly allocated to that job. This prevents users from accessing GPUs beyond their allocation and solves the resource contention issue. Increasing memory allocation (option A) does not restrict GPU access, increasing priority (option C) only affects scheduling order, and adding CPU requests (option D) does not control GPU visibility.
Why the other options are wrong
- A. Increasing memory allocation per job does not restrict access to unallocated GPUs; it only modifies memory resources.
- C. Setting higher priority only affects job scheduling order, not device access restrictions within running jobs.
- D. Adding CPU core requests to job scripts does not prevent users from accessing GPUs outside their allocation.
Question 9
A new researcher needs access to GPU resources but should not have permission to modify cluster settings or manage other users.
What role should you assign them in Run:ai?
Show answer and explanation
Correct answer: A. L1 Researcher
L1 Researcher is the appropriate role for a new researcher who needs GPU resource access without administrative privileges. This role grants access to GPU clusters and job submission capabilities while restricting permissions to modify cluster settings or manage other users. Department Administrator and Application Administrator roles carry administrative responsibilities beyond what a researcher requires, while Research Manager involves user management duties.
Why the other options are wrong
- B. Department Administrator has permissions to manage cluster settings and user accounts, exceeding the researcher's needs.
- C. Application Administrator manages application-level configurations and cluster administration, not appropriate for a basic researcher.
- D. Research Manager typically involves managing other researchers and team resources, which conflicts with the requirement to not manage other users.
Question 10
When troubleshooting Slurm job scheduling issues, a common source of problems is jobs getting stuck in a pending state indefinitely.
Which Slurm command can be used to view detailed information about all pending jobs and identify the cause of the delay?
Show answer and explanation
Correct answer: A. scontrol
The scontrol command provides detailed diagnostic information about job states and reasons for delays, including why jobs remain pending. It can display comprehensive job configuration and status details that help identify scheduling obstacles. While sacct shows historical job accounting data and sinfo displays node/partition information, neither provides the detailed real-time job state information needed to diagnose pending job issues.
Why the other options are wrong
- B. sacct displays historical job accounting information after jobs complete, not real-time status of pending jobs.
- C. sinfo shows partition and node status, but does not provide detailed job-specific information about why individual jobs are pending.
That was 10 of 66.
The full NVIDIA NCP-AIO pack has all 66 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.