NCP-AII Question 49
Select 3You are configuring network fabric ports for a high-performance AI cluster using NVIDIA networking solutions. The cluster requires optimal connectivity between GPU nodes and storage servers over RDMA (Remote Direct Memory Access). Which of the following configurations should you implement to ensure low-latency, high-throughput communication?
- A
Enable RDMA over Converged Ethernet (RoCE) on the network switches and hosts.
- B
Configure Quality of Service (QoS) to prioritize RDMA traffic in the network fabric.
- C
Disable flow control mechanisms on all ports to avoid packet delays.
- D
Ensure that all network ports are using the correct MTU size (e.g., 9000 for jumbo frames).
- E
Use Spanning Tree Protocol (STP) to manage redundancy in the network fabric.
Show answer and explanation
Correct answers: A, B, D
Explanation
To configure network fabric ports for hosts in a high-performance AI cluster, it is essential to enable RDMA over Converged Ethernet (RoCE) for low-latency communication, configure QoS to prioritize RDMA traffic, and ensure the correct MTU size (e.g., jumbo frames) is set for efficiency. Disabling flow control would degrade performance, and Spanning Tree Protocol (STP) is not suitable for such use cases due to its limitations in high-performance environments.
- A. Correct.
RDMA over Converged Ethernet (RoCE) is crucial for achieving low-latency, high-throughput communication in AI workloads by enabling direct memory access between nodes without involving the host CPU.
- B. Correct.
Configuring Quality of Service (QoS) ensures that RDMA traffic is prioritized, reducing latency and improving performance for critical workloads.
- C. Incorrect.
Disabling flow control mechanisms can lead to packet loss and increased latency, which negatively impacts RDMA performance. Flow control should remain enabled to ensure efficient communication.
- D. Correct.
Using the correct MTU size, such as 9000 for jumbo frames, reduces overhead and improves efficiency in high-performance networking scenarios like RDMA.
- E. Incorrect.
Spanning Tree Protocol (STP) is typically not used in high-performance AI clusters due to its potential to introduce latency and limit bandwidth utilization. Instead, other redundancy protocols like MLAG or EVPN are preferred.