NCP-AII Question 33
Select 3You are deploying a new AI workload on a GPU-accelerated server in a data center. During testing, you notice that the workload is not performing as expected, with significant delays in data transfer between the server and the storage system. Which of the following actions would help identify and resolve the issue?
- A
Check the server's PCIe bandwidth utilization to ensure the GPUs are not limited by data transfer rates.
- B
Verify that the storage system supports NVMe over Fabrics (NVMe-oF) for high-speed data transfer.
- C
Upgrade the server's CPU to a higher core count to match the GPU processing requirements.
- D
Ensure that the network interface card (NIC) on the server supports the required bandwidth for the workload.
- E
Inspect the server's power supply to confirm it meets the recommended wattage for the GPUs.
Show answer and explanation
Correct answers: A, B, D
Explanation
AI workloads require optimized systems and networks to perform efficiently. Bottlenecks in PCIe bandwidth, storage system speed, or network bandwidth can significantly impact performance. By analyzing PCIe utilization, confirming NVMe-oF support, and ensuring the NIC has sufficient bandwidth, you can identify and resolve the data transfer delays. CPU upgrades or power supply checks, while important for other scenarios, do not directly address the issue described in this scenario.
- A. Correct.
Checking the server's PCIe bandwidth utilization is essential because GPUs heavily rely on PCIe lanes for data transfer. Insufficient bandwidth can lead to bottlenecks.
- B. Correct.
Verifying the storage system's support for NVMe-oF is relevant because it ensures faster data transfer rates, which are crucial for AI workloads that process large datasets.
- C. Incorrect.
Upgrading the CPU may not directly address the data transfer issue. While CPU-GPU compatibility is important, the problem here is network and storage latency.
- D. Correct.
The NIC's bandwidth plays a critical role in data transfer between the server and external systems. Ensuring it meets workload requirements can resolve network-related delays.
- E. Incorrect.
Inspecting the power supply is important for stability, but it is unrelated to the specific issue of data transfer delays between the server and storage.