試験問題と解答はNCP-AII学習ガイド問題を試そう！ [Q19-Q40]

試験問題と解答はNCP-AII学習ガイド問題を試そう！

NVIDIA AI Infrastructure認証サンプル問題と練習試験合格させます

質問 # 19
Which of the following are key benefits of using NVIDIA Spectrum-X switches in an A1 infrastructure compared to traditional Ethernet switches? (Select THREE)

A. Native support for IPv6.
B. Support for RoCE (RDMA over Converged Ethernet) and InfiniBand protocols, enabling high-bandwidth, low-latency communication.
C. Advanced telemetry and monitoring capabilities for network performance optimization.
D. Lower cost per port.
E. Hardware-based acceleration for collective communication operations used in distributed A1 training.

正解：B、C、E

解説：
Spectrum-X switches are designed for high-performance computing and A1 workloads. They support RoCE and InfiniBand for low- latency communication, offer advanced telemetry for network optimization, and include hardware-based acceleration for collective communication operations, improving the efficiency of distributed A1 training. While Spectrum-X supports IPv6, this is also a common feature in modern Ethernet switches. Spectrum-X switches typically have a higher cost per port compared to basic Ethernet switches due to their advanced features and performance.

質問 # 20
A data scientist reports slow data loading times when training a large language model. The data is stored in a Ceph cluster. You suspect the client-side caching is not properly configured. Which Ceph configuration parameter(s) should you investigate and potentially adjust to improve data loading performance? Select all that apply.

A. fuse_client_max_background
B. client quota
C. client cache size
D. mds cache size

正解：A、C

解説：
Client-side caching in Ceph is primarily controlled by 'client cache size' which determines the amount of memory the Ceph client uses for caching data. 'mds cache size' controls the metadata server cache size, impacting metadata operations. controls the maximum number of background requests a FUSE client can make, influencing concurrency. affects the number of threads used by the OSDs, not the client-side caching, and 'client quota' limits storage usage, not caching.

質問 # 21
You're experiencing unexpected memory errors during the execution of a large-scale deep learning training job. Suspecting faulty memory modules, which diagnostic tool is best suited to perform a comprehensive memory test on the server's RAM?

A. Ishw
B. nvprof
C. nvidia-smi
D. memtest86+
E. DCGM

正解：D

解説：
'memtest86+' is specifically designed for performing thorough memory tests on system RAM, identifying potential errors and hardware faults. 'nvidia-smi' is for GPU monitoring, 'nvprof is a profiler, DCGM can test GPU memory, but not system RAM specifically. 'Ishw' lists hardware information, but doesn't perform diagnostic tests.

質問 # 22
Your AI training pipeline involves a pre-processing step that reads data from a large HDF5 file. You notice significant delays during this step. You suspect the HDF5 file structure might be contributing to the slow read times. What optimization technique is MOST likely to improve read performance from this HDF5 file?

A. Converting the HDF5 file to a CSV file.
B. Compressing the HDF5 file using gzip.
C. Storing the HDF5 file on a network file system like NFS.
D. Encrypting the HDF5 file for enhanced security.
E. Reorganizing the HDF5 file to improve data contiguity and chunking.

正解：E

解説：
Reorganizing the HDF5 file (option C) to improve data contiguity and chunking is the most effective optimization. HDF5 performance is highly dependent on how the data is laid out within the file. Contiguous data and optimal chunk sizes allow for more efficient 1/0 operations. Converting to CSV (A) loses the hierarchical structure of HDF5. Storing on NFS (B) adds network overhead. Compression (D) can reduce storage space but increases decompression overhead. Encryption (E) adds overhead without improving read performance.

質問 # 23
You encounter a situation where an NVIDIA driver installation fails with the error message 'ERROR: Unable to load the kernel module 'nvidia.ko'. This may be because it was built for another kernel...'. Assuming the kernel headers are correctly installed, what is the most likely cause and solution?

A. Secure Boot is enabled, and the NVIDIA kernel module is not signed. Solution: Sign the NVIDIA kernel module with a machine owner key (MOK).
B. The DKMS module build failed. Solution: Manually rebuild the DKMS module using 'dkms autoinstall'.
C. The 'nouveau' driver is still loaded. Solution: Blacklist 'nouveau- and reboot.
D. The system is running out of disk space in '/tmp'. Solution: Free up space or remount S/tmp' with more space.
E. The NVIDIA driver version is incompatible with the current kernel version. Solution: Install a compatible driver version.

正解：A、B、E

解説：
The error message indicates a mismatch between the driver and kernel. This can happen due to incompatibility, unsigned modules due to Secure Boot, or a failed DKMS build. Installing a compatible version, MOK signing, and rebuilding DKMS are valid solutions. While nouveau can interfere, the error message points specifically to a module loading problem, making the other options more likely. Lack of disk space is a less common, but possible, issue.

質問 # 24
You are experiencing link flapping (frequent up/down transitions) on several InfiniBand links in your AI infrastructure. This is causing intermittent connectivity issues and performance degradation. What are the MOST likely causes of this issue, and what steps should you take to troubleshoot and resolve it? (Select TWO)

A. Excessive broadcast traffic causing congestion.
B. Mismatched link speeds or duplex settings between connected devices.
C. Incorrect MTU (Maximum Transmission Unit) configuration on the affected interfaces.
D. Software bugs in the operating system or InfiniBand drivers.
E. Faulty or damaged cables, connectors, or transceivers.

正解：B、E

解説：
Link flapping is most commonly caused by physical layer issues (faulty cables, connectors, or transceivers) or configuration mismatches (link speeds or duplex settings). Troubleshooting should focus on inspecting the physical connections and verifying that the link speed and duplex settings are correctly configured on both ends of the link. While MTIJ issues and software bugs can cause network problems, they are less likely to directly cause link flapping. Excessive broadcast traffic can cause performance issues but is less likely to result in frequent link up/down transitions.

質問 # 25
You are designing a network for a distributed training job utilizing multiple GPUs across multiple nodes. Which network characteristic is MOST critical for minimizing training time?

A. Large MTU
B. Low cost
C. High packet loss rate
D. Low latency
E. High bandwidth

正解：A、D、E

解説：
Distributed training involves frequent data exchange between GPUs. High bandwidth is necessary to move large datasets quickly. Low latency reduces the round-trip time for communication. Large MTIJ (Maximum Transmission Unit) can reduce overhead by transmitting larger packets. High packet loss is detrimental to performance. Low cost should not be prioritized over performance in a distributed training environment.

質問 # 26
You've flashed the BlueField OS to your SmartNlC, but you need to customize the kernel command line arguments (bootargs) to enable a specific feature. Where is the MOST appropriate place to modify these arguments for persistent changes that survive reboots?

A. Passing it as an argument to bfboot during deployment.
B. In the '/proc/cmdline' file. This allows immediate changes.
C. In the bootloader configuration file (e.g., extlinux.conf or grub.cfg) on the BlueFieId's flash memory.
D. Directly in the kernel image file itself using a hex editor.
E. In the '/etc/default/grub' file on the BlueField OS, followed by updating the GRUB configuration.

正解：C

解説：
The bootloader configuration file (extlinux.conf, grub.cfg, uEnv.txt depending on the system) is where boot arguments are persistently stored. Modifying the kernel image directly is highly discouraged and risky. 'letc/default/grub' is a common location on standard Linux systems, but not necessarily on the BlueField OS's boot environment. '/proc/cmdline' shows the currently used arguments, but modifying it doesn't persist changes across reboots. bfboot will only change the image during that flash, changes at the bootloader level persist after subsequent flashes.

質問 # 27
You have a GPU-intensive application that requires the latest features of CUDA 12. However, your host system's NVIDIA driver is only compatible with CUDA 11.8. What steps can you take to enable your application to use CUDA 12 within a Docker container, without upgrading the host driver?

A. Mount the CUDA 12 libraries from a separate Docker volume into the container and configure the accordingly.
B. Downgrade the application to use CUDA 11.8 to match the host's driver version.
C. Use a Docker image based on 'nvidia/cuda:12.0-base-ubuntu20.04'. The NVIDIA Container Toolkit will automatically handle the driver compatibility between the host and the container.
D. Upgrade the NVIDIA driver on the host system to the latest version compatible with CUDA 12.
E. Install CUDA 12 inside the Docker container and set the 'CUDA DRIVER VERSION' environment variable to match the host driver version.

正解：C

解説：
The NVIDIA Container Toolkit enables compatibility between the host driver and the CUDA version within the container (B). Using a Docker image based on will allow your application to leverage CUDA 12 features. Upgrading the host driver (A) is an option but not necessary and may introduce other compatibility issues. Setting (C) is not a standard or reliable approach. Mounting CUDA libraries from a volume (D) is complex and might not resolve driver version mismatches. Downgrading the application (E) avoids the problem but sacrifices access to CUDA 12 features. Because NVIDIA ensures a degree of backwards compatibility, the newer toolkit in the container can often work with an older driver on the host.

質問 # 28
A team is validating a DGX BasePOD deployment. Using cmsh, they run a command to check GPU health across all nodes. What indicates that the system is ready for AI workloads?

A. All GPUs report Status_Health = OK and Health = OK for each device.
B. At least half of the GPUs report Status_Health = OK.
C. Only the head node's GPUs need to be healthy.
D. The command output is ignored if the system powers on without errors.

正解：A

解説：
In an NVIDIA DGX BasePOD or SuperPOD environment, "Cluster Health" is a binary state: either the entire fabric and all compute resources are ready, or the cluster is considered degraded. Using the Bright Cluster Manager (BCM) shell (cmsh), administrators can aggregate telemetry from every node in the cluster. For a system to be considered "Production Ready," every single GPU across the multi-node deployment must report a status of Health = OK. This verification ensures that the hardware is communicating correctly over the PCIe bus, the NVLink fabric is initialized, and no ECC (Error Correction Code) memory errors are present. If even a single GPU in a 32-node cluster is unhealthy, collective communication libraries like NCCL may hang or experience significant performance penalties during "All-Reduce" operations, as the entire job typically scales to the speed of the slowest/unhealthiest component. Therefore, seeing Status_Health = OK for every device is the mandatory exit criterion for the bring-up phase.

質問 # 29
An AI inference workload running on a server equipped with NVIDIA TensorRT is experiencing intermittent performance degradation. Profiling reveals potential issues with GPU utilization. Which command-line tool is best suited to provide real-time GPU performance metrics, including utilization, memory usage, and power consumption, during workload execution?

A. nvprof
B. nvidia-smi
C. nvitop
D. gpustat
E. tegrastats

正解：B

解説：
'nvidia-smi' (NVIDIA System Management Interface) is the primary tool for monitoring GPU performance in real-time, including utilization, memory usage, and power consumption. While 'nvprof is a profiler, it's more suited for in-depth analysis after the run. 'nvitop' and gpustat are third-party tools that provide similar functionalities as -nvidia-smr. 'tegrastats' is specific to NVIDIA Jetson devices.

質問 # 30
A system administrator receives an alert about a potential hardware fault on an NVIDIA DGX A100. The GPU performance seems degraded, and the system fans are operating loudly. What step should be recommended to identify and troubleshoot the hardware fault?

A. Power drain then restart the DGX and check if the performance degradation resolves.
B. Run a deep learning workload to stress test the GPUs and check whether the issue persists.
C. Check the NVIDIA System Management Interface (nvidia-smi) for GPU status and temperatures.
D. Increase the fan speed to maximum and check whether the performance improves.

正解：C

解説：
When a DGX system exhibits high fan speeds and performance degradation, it is typically engaging in Thermal Throttling. High-performance GPUs like the A100 or H100 will automatically reduce their clock speeds (and thus performance) if they exceed safe temperature thresholds. The first and most critical diagnostic step is to run nvidia-smi. This utility provides immediate, real-time telemetry on GPU temperatures, power draw, and "Clocks Throttle Reasons." By reviewing the output, an administrator can see if "Thermal" is listed as the reason for reduced clocks. This identifies whether the issue is environmental (blocked airflow/hot aisle temperature) or hardware-specific (a failed GPU thermal interface or a dead internal fan). Running more workloads (Option A) would exacerbate the heat, while a power drain (Option C) is a
"last resort" that doesn't provide diagnostic data. nvidia-smi provides the evidentiary data needed to determine if an RMA (Return Merchandise Authorization) is required for the GPU tray.

質問 # 31
A media company is developing an AI platform for video content analysis that requires storing and processing large volumes of unstructured video data. The platform must support high throughput for data ingestion and provide efficient access for real-time analytics. Given these requirements, which storage strategy should the company implement?

A. Tape storage for its cost-effectiveness and archival capabilities
B. File storage for hierarchical organization and easy navigation
C. Block storage for low latency and high performance
D. Object storage for scalability and metadata management

正解：B

解説：
While object storage is excellent for massive scale and metadata, NVIDIA AI infrastructure best practices for training workloads-especially video analysis-heavily prioritizeParallel File Systems (PFS). Modern AI frameworks (PyTorch, TensorFlow) and NVIDIA's own SDKs (like DeepStream or NeMo) are optimized to read from POSIX-compliant file systems. For video content analysis, the training process involves "sharding" large video files and performing random-access reads across a massive dataset. A high-performance file system (such as Lustre, Weka, or IBM Storage Scale) provides the high throughput and low-latency metadata operations required to keep 8 or more H100 GPUs per node saturated with data. File storage allows for the hierarchical organization that data scientists use to manage datasets (e.g., /datasets/train/videos/) and supports GPUDirect Storage (GDS), which allows the GPU to pull data directly from the storage fabric into GPU memory, bypassing the CPU to maximize ingestion throughput.

質問 # 32
Which of the following are key considerations when choosing between CPU pinning and NUMA (Non-Uniform Memory Access) awareness for a distributed training job on a multi-socket AMD EPYC server with multiple GPUs?

A. CPU pinning ensures that each process/thread runs on a specific CPU core, reducing context switching overhead. NUMA awareness ensures that the CPU cores and memory used by a process are located within the same NUMA node, minimizing memory access latency.
B. NUMA awareness is generally more important than CPU pinning because it directly impacts memory bandwidth.
C. Neither CPU pinning nor NUMA awareness are relevant for GPIJ-accelerated workloads, as the GPUs handle all the computation.
D. CPU pinning is generally more important than NIJMA awareness because it directly impacts CPU utilization.
E. Both CPU pinning and NUMA awareness are critical for optimizing performance. They should be used in conjunction to achieve optimal performance.

正解：A、E

解説：
CPU pinning and NUMA awareness are both important for optimizing performance on multi-socket systems. CPU pinning reduces context switching overhead, while NIJMA awareness reduces memory access latency. Using them together provides the best performance. They ensure the CPU and memory being used are on the same node, which reduces latency. While GPUs handle intense computation, CPUs prepare and feed data to the GPUs, and are critical for performance.

質問 # 33
You encounter a situation where a container running with GPU support is experiencing significant performance degradation compared to running the same application directly on the host. You have already verified that the NVIDIA drivers are correctly installed and the NVIDIA Container Toolkit is properly configured. Which of the following could be contributing factors to this performance difference?
(Select all that apply)

A. The container is using a significantly older version of the CUDA runtime compared to the host.
B. CPU pinning or NIJMA affinity is not properly configured for the container, leading to inefficient memory access.
C. Insufficient bandwidth between CPU and GPU
D. The kernel version within the container is significantly different from the host kernel, leading to driver compatibility issues.
E. The '-ipc=host' flag is not used when running the container, causing inter-process communication overhead.

正解：A、B

解説：
Using an older CUDA runtime within the container (A) can lead to performance degradation due to missing optimizations or compatibility issues with the application. Improper CPU pinning and NUMA affinity (B) can cause the container to access memory inefficiently, especially in multi-socket systems. '--ipc=host' (C) can improve performance in some cases by sharing the host's IPC namespace, but it's not always necessary and can have security implications. Kernel version differences (D) are generally handled by the NVIDIA Container Toolkit, which ensures compatibility. Insufficient bandwidth between CPU and GPU (E) might be caused by hardware issue.

質問 # 34
Consider a scenario where you're using GPUDirect Storage to enable direct memory access between GPUs and NVMe drives. You observe that while GPUDirect Storage is enabled, you're not seeing the expected performance gains. What are potential reasons and configurations you should check to ensure optimal GPUDirect Storage performance? Select all that apply.

A. Verify that the NVMe drives are properly configured in a RAID 0 configuration.
B. Disable CPU-side caching to force all I/O operations to go directly to the GPU memory.
C. Ensure that the NVMe drives are connected to the system via PCle Gen4 or Gen5.
D. Confirm that the CUDA driver version is compatible with GPIJDirect Storage.
E. Check if the file system supports direct I/O (e.g., using 'directio' mount option).

正解：C、D、E

解説：
Explanation:GPUDirect Storage requires PCle Gen4/Gen5 for sufficient bandwidth (B). The CUDA driver must be compatible with GPUDirect Storage (C). Direct I/O support in the file system is essential to bypass the OS cache and allow direct GPU access (D). RAID 0 (A) is about storage speed but not directly related to GDS functionality. Disabling CPU-side caching (E) is usually detrimental as it can reduce overall system performance. Note, this is not always bad but needs to be tested depending on application.

質問 # 35
You are tasked with creating a custom Docker image for a deep learning application that requires a specific version of cuDNN. You want to minimize the image size while ensuring that the cuDNN libraries are correctly installed and configured. What is the most efficient way to achieve this?

A. Use the NVIDIA Container Toolkit to dynamically inject the cuDNN libraries into the container at runtime.
B. Use a pre-built CUDA base image and install cuDNN during the container run.
C. Download the cuDNN archive from NVIDIA, extract the libraries, and manually copy them to the appropriate locations within the Dockerfile.
D. Use a multi-stage Docker build, using a base image with the desired CUDA version for building and then copying only the necessary cuDNN libraries to a smaller runtime image.
E. Install the entire CUDA toolkit within the Docker image, even if only cuDNN is needed.

正解：D

解説：
A multi-stage Docker build (B) is the most efficient approach. It allows you to use a larger image with the CUDA toolkit for building and then copy only the necessary cuDNN libraries to a smaller runtime image, minimizing the final image size. Manually copying libraries (A) is tedious and error-prone. Installing the entire CUDA toolkit (C) unnecessarily increases the image size. The NVIDIA Container Toolkit (D) focuses on enabling GPU access, not dynamically injecting specific libraries. Running an install of cuDNN during the container run is problematic since the image should be self-contained.

質問 # 36
You are configuring a BlueField-3 DPLJ for a cloud-native application using Kubernetes. You want to offload container networking using OVS (Open vSwitch). Which of the following configuration steps are NECESSARY to integrate the BlueField-3 DPIJ with the Kubernetes cluster for network offload? (Select TWO)

A. Deploy the NVIDIA BlueField Kubernetes Operator to manage the DPIJ lifecycle and networking configurations.
B. Configure the Kubernetes CNI (Container Network Interface) to use the OVS bridge managed by the BlueField DPIJ.
C. Manually create OVS bridges and vPorts on the BlueField DPU using 'ovs-vsctr.
D. Install the Mellanox OFED drivers on all Kubernetes worker nodes.
E. Configure the BlueField DPIJ to act as a DHCP server for the Kubernetes pods.

正解：A、B

解説：
The NVIDIA BlueField Kubernetes Operator is essential for automating the management and configuration of the DPIJ within the Kubernetes environment. This includes creating and managing OVS bridges. Integrating the Kubernetes CNI to use the OVS bridge managed by the BlueField DPIJ allows pod networking traffic to be offloaded to the DPU. Installing Mellanox OFED everywhere isn't needed with the operator. While you could manually create the bridges (E), the operator is the preferred method. The DPIJ acting as a DHCP server (D) is not a requirement for simple network offload.

質問 # 37
After installing a new NVIDIA GPU, you attempt to run a CUDA application, but you encounter the following error: 'CUDA error: CUDA driver version is insufficient for CUDA runtime version'. You have verified the driver and CUDA toolkit are installed. What is the MOST likely reason for this error, and how do you resolve it?

A. The NVIDIA driver is too old for the CUDA toolkit. Update the NVIDIA driver to a version that supports the CUDA toolkit.
B. The CUDA toolkit is too old. Update the CIJDA toolkit.
C. The CUDA runtime libraries are missing from the system path. Add them to the PATH variable.
D. The GPU is not compatible with the CUDA toolkit. Install a different GPIJ.
E. The CUDA VISIBLE DEVICES environment variable is not set correctly.

正解：A

解説：
This error indicates an incompatibility between the driver and the CUDA toolkit. The most common reason is an outdated driver. The driver must be at least as new as the CUDA toolkit's minimum required driver version. CUDA VISIBLE DEVICES relates to GPU selection, not driver version.

質問 # 38
After upgrading the network card drivers on your A1 inference server, you experience intermittent network connectivity issues, including packet loss and high latency. You've verified that the physical connections are secure. Which of the following steps would be most effective in troubleshooting this issue?

A. Reinstall the operating system.
B. Roll back the network card drivers to the previous version.
C. Update the server's BIOS.
D. Run network diagnostic tools like 'ping', 'traceroute', and 'iperf3' to assess the network performance.
E. Check the system logs for error messages related to the network card or driver.

正解：B、D、E

解説：
Rolling back drivers is a quick way to revert to a known working state. Checking system logs will provide valuable information about driver errors or network issues. Network diagnostic tools will quantify the network performance and help isolate the problem. Reinstalling the OS is drastic and should be a last resort. Updating the BIOS is unlikely to resolve driver-related network issues unless specifically recommended for the network card.

質問 # 39
Which of the following are valid methods for verifying the health and connectivity of InfiniBand links in an NCP-AII environment? (Select TWO)

A. Using 'ping' to test basic IP connectivity over the InfiniBand interface.
B. Checking the system logs ( ' /var/log/messages' or equivalent) for any InfiniBand-related error messages.
C. Using 'netstat' to check TCP connections.
D. Using 'ibstat' to check the link state, physical state, and other relevant parameters of InfiniBand ports.
E. Using 'sminfo' to query the Subnet Manager for network topology and status information.

正解：D、E

解説：
'ibstat' is a command-line utility specifically designed for checking the status of InfiniBand ports. 'sminfo' allows you to communicate with the Subnet Manager and retrieve network topology and status. While 'ping' can verify IP connectivity over InfiniBand, it doesn't directly assess the health of the InfiniBand link itself. Checking system logs is a useful supplementary task but isn't the primary method.

質問 # 40
......

NVIDIA NCP-AII 認定試験の出題範囲：

トピック	出題範囲
トピック 1	トラブルシューティングと最適化: GPU、ネットワークカード、電源などの障害のあるハードウェアコンポーネントの特定と交換、および AMD Intel サーバーとストレージのパフォーマンスの最適化について説明します。
トピック 2	クラスターのテストと検証: HPL および NCCL ベンチマーク、NVLink およびファブリック帯域幅テスト、ケーブルおよびファームウェアのチェック、HPL、NCCL、および NeMo を使用したバーンインテストによる完全なクラスター検証をカバーします。
トピック 3	物理層管理: BlueField ネットワークプラットフォームデバイスの構成と、AI および HPC ワークロード用のマルチインスタンス GPU (MIG) パーティショニングの設定について説明します。
トピック 4	システムとサーバーの立ち上げ: BMC OOB TPM 構成、ファームウェアのアップグレード、ハードウェアのインストール、電源と冷却の検証など、GPU ベースの AI インフラストラクチャのエンドツーエンドの物理セットアップをカバーし、サーバーがワークロードに対応できる状態であることを確認します。
トピック 5	コントロールプレーンのインストールと構成: Base Command Manager、OS、Slurm Enroot Pyxis、NVIDIA GPU および DOCA ドライバー、コンテナーツールキット、NGC CLI を含むソフトウェアスタックの展開について説明します。