Q19. A Docker container running a CUDA application terminates unexpectedly with an ‘out of memory’ error, despite the host machine having sufficient RAM. What are the potential causes and how would you diagnose them?

Q20. You’ve deployed a container from NGC containing a computationally intensive AI model training script. You notice that the container is consistently being killed by the Kubernetes OOMKiller, even though the node has sufficient memory available. What are the possible causes and solutions?

Q21. You’re deploying a DOCA-based firewall application on a BlueField-2 DPU. The application uses eBPF for packet filtering. What is the primary reason for using eBPF in this scenario?

Q22. Consider a scenario where you’re trying to run a Docker container that uses the NVIDIA MPS (Multi-Process Service). However, you keep encountering errors indicating that MPS is not properly initialized within the container. What steps should you take to troubleshoot this issue?

Q23. Consider the following Kubernetes pod definition:

What does the ‘nvidia.com/gpu: 1’ setting achieve?

Q24. You want to upgrade the NVIDIA drivers on your Kubernetes nodes without disrupting the running AI workloads. What is the recommended approach to perform a rolling upgrade of the NVIDIA drivers?

Q25. You are troubleshooting an issue where a container inside a pod is unable to access the NVIDIA GPU. The NVIDIA Device Plugin is running, and the pod is requesting ‘nvidia.com/gpu: 1’. What are the potential causes for this issue?

Q26. You are using NVSHMEM to manage shared memory across multiple GPUs in a multi-node cluster. Your application is crashing with out- of-memory errors, even though the reported GPU memory usage is well below the total available. You have already confirmed sufficient physical RAM on all nodes. What is the MOST likely cause, related to NVSHMEM configuration, of these out-of-memory errors?

Q27. You are deploying a stateful application to your Kubernetes cluster running on NVIDIA hardware provisioned through BCM. This application requires direct access to a persistent volume on a high-performance NVMe drive. Which of the following methods is MOST appropriate for providing this access while ensuring high performance and data consistency?

Q28. You have a Slurm cluster configured with multiple partitions, and you want to restrict a specific user group to only submit jobs to a particular partition. How can you achieve this using Slurm’s Quality of Service (QOS) and Access Control features?

Q29. A system administrator is looking to set up virtual machines in an HGX environment with NVIDIA Fabric Manager.
What three (3) tasks will Fabric Manager accomplish? (Choose three.)

Q30. You want to configure a Slurm cluster with heterogeneous nodes, some equipped with high-performance GPUs and others with only CPUs. Which Slurm configuration parameter allows you to define and utilize these different node types effectively?

Q31. A data scientist submits a Run.ai job requesting 4 GPUs. However, due to resource constraints, only 2 GPUs are immediately available. You want the job to automatically start running as soon as the remaining 2 GPUs become available, without manual intervention. How do you configure Run.ai to achieve this?

Q32. You are deploying an AI application using Fleet Command. You want to ensure that the application automatically restarts if it crashes on an edge device. How can you achieve this?

Q33. Consider this YAML snippet for deploying the NVIDIA device plugin. Which statement is true about the highlighted segment?

Q34. You need to configure node health checks in Slurm to automatically detect and drain unhealthy nodes. Which of the following approaches is the MOST robust and recommended way to achieve this?

Q35. You’re deploying an AI inference application using NVIDIA Triton Inference Server in a Kubernetes cluster. Which of the following storage options is MOST suitable for storing the trained models, considering scalability and access speed?

Q36. You’re configuring MIG on an NVIDIAA100 for a mixed AI/HPC environment. One application requires high memory bandwidth, and the other requires high compute throughput. Which MIG instance configuration would optimally balance these requirements?

Q37. A data science team is using Fleet Command to deploy AI models to edge devices in a smart city project. They’ve noticed that some devices are consistently failing to update due to insufficient disk space. Which of the following is the MOST effective strategy to mitigate this issue?

Q38. 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?

Q39. In a high availability (HA) cluster, you need to ensure that split-brain scenarios are avoided.
What is a common technique used to prevent split-brain in an HA cluster?

Q40. You are designing a data center network to support distributed deep learning training across multiple servers. The training job uses NCCL (NVIDIA Collective Communications Library) for inter-GPU communication. Which of the following network configurations will maximize the performance of NCCL?