4.3 Describe High-Performance Network Enabling Technologies for AI in Data Center Infrastructure 350-601 Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   What is the purpose of Explicit Congestion Notification (ECN) in a high-performance network?

   • ATo synchronize clocks between switches.
   • BTo provide feedback about congestion without dropping packets.
   • CTo prioritize traffic based on VLAN tags.
   • DTo encrypt data in transit.
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A network team is deploying a RoCEv2 fabric for AI training and wants to use Data Center TCP (DCTCP) to handle congestion. The team needs to ensure that the switches mark packets with ECN when congestion is detected, and that the end hosts react by reducing their sending rate. Which configuration is required on the switches?

   • AEnable ETS to prioritize the RoCE class
   • BEnable ECN marking with a threshold and configure the RoCE class to use ECT(0)
   • CEnable PFC with a low threshold to prevent marking
   • DDisable ECN and rely on PFC only
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A research lab is running distributed training using Horovod over a RoCEv2 fabric. The lab's network uses a spine-leaf topology with 4 spines and 8 leaves. The team notices that the training throughput is lower than expected. They suspect that Horovod's communication pattern is not optimized for the network topology, causing excessive traffic on the spine links. Which action should they take to improve performance?

   • AIncrease the number of spines from 4 to 8
   • BEnable PFC on all traffic classes
   • CConfigure Horovod to use the hierarchical allreduce algorithm, which reduces cross-spine traffic by aggregating within a rack first
   • DDisable ECN marking on the RoCE class
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A data science team is running distributed training using NVIDIA NCCL (NVIDIA Collective Communications Library) over a RoCEv2 fabric. The training performance is poor because NCCL's default communication pattern is not optimized for the network topology. The team wants to improve performance by making NCCL aware of the network topology so it can schedule communications more efficiently. What should they do?

   • AEnable ECN marking on all traffic to reduce congestion
   • BIncrease the PFC pause threshold on all switches to reduce pauses
   • CUse a single TCP socket for all NCCL communication
   • DEnable NCCL's topology-aware communication by setting the NCCL_TOPO_FILE environment variable to describe the network hierarchy
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   A network engineer is configuring PFC on a switch that carries both RoCEv2 and regular TCP traffic. The engineer is concerned that PFC on the RoCE class could cause head-of-line blocking for the TCP traffic. Which configuration best mitigates this risk?

   • APlace RoCEv2 in a dedicated PFC priority class and TCP in a separate non-PFC class
   • BEnable PFC on all classes to ensure fairness
   • CDisable PFC entirely and rely on TCP retransmission for RoCE
   • DPlace RoCEv2 and TCP traffic in the same PFC priority class
   Reveal Solution Discussion Select an answer first