NCP-AII

NCP-AII practice questions, page 4 of 7, with answer review and explanations.

Question 16

To validate bisectional bandwidth across two racks in a Spectrum-X Ethernet fabric, which NCCL test configuration isolates East-West traffic?

  1. Run without splits and analyze per-rack averages
  2. NCCL_TESTS_SPLIT= “MOD 2” ./all_reduce_perf -g 8
  3. NCCL_TESTS_SPLIT= “DIV 8” ./all_reduce_perf -g 1
  4. NCCL_TESTS_SPLIT= “OR 0x7” ./all_reduce_perf -g 8

Question 17

A network engineer is tasked with configuring the management, storage, and compute networks for a new DGX BasePOD deployment. Which statement best describes the network segmentation required for optimal operation?

  1. A single VLAN for all types of network traffic.
  2. Two networks: one for management and one for compute.
  3. Four networks: compute, storage, out-of-band, and management.

Question 18

During cluster deployment, the UFM Cable Validation Tool reports “Wrong-neighbor” errors on multiple InfiniBand links. What is the most efficient way to resolve this issue?

  1. Verify LLDP data against topology files and remediate.
  2. Reboot all leaf switches to force LLDP rediscovery.
  3. Disable FEC on all switches to bypass neighbor validation.
  4. Replace all affected cables with higher-grade OM5 fiber optics.

Question 19

During HPL execution on a DGX cluster, the benchmark fails with “not enough memory” errors despite sufficient physical RAM. Which HPL.dat parameter adjustment is most effective?

  1. Increase block size to 6144 to maximize GPU utilization.
  2. Reduce the problem size while maintaining the same block size.
  3. Set PMAP to 1 to enable process mapping.
  4. Disable double-buffering via BCAST parameter.

Question 20

Your company is planning to expand its AI capabilities significantly over the next five years. To future-proof your storage infrastructure, you need a solution that can scale in both capacity and performance. Which of the following strategies best ensures that your storage infrastructure remains adaptable to future AI demands?

  1. Implement on-premises block storage system with periodic hardware upgrades.
  2. Implement single-tier cloud storage solution to leverage cloud scalability.
  3. Use a hybrid cloud model combining scalable cloud resources with on-premises infrastructure.
  4. Deploy an all-flash array and remove data tiering to reduce latency.