4.3 Describe High-Performance Network Enabling Technologies for AI in Data Center Infrastructure 350-601 Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A network team is tuning a lossless Ethernet fabric for an AI cluster. They observe that during heavy training epochs, switch buffers fill up and PFC pause frames are triggered frequently, which causes the RoCE traffic to stall and reduces overall GPU utilization. The team wants to reduce the frequency of PFC pauses by signaling congestion earlier so that the end-to-end congestion control mechanism can react before buffers overflow. Which feature should they enable on the switches?

   • AEnhanced Transmission Selection (ETS) with a higher bandwidth allocation for RoCE
   • BPriority Flow Control (PFC) with a higher pause threshold
   • CData Center TCP (DCTCP) disabled on the RoCE class
   • DExplicit Congestion Notification (ECN) with a marking threshold below the PFC pause threshold
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   In a PFC-enabled network, what happens when a switch buffer for a specific priority class reaches its threshold?

   • AThe switch marks packets with ECN.
   • BThe switch sends a pause frame to the upstream device for that priority class.
   • CThe switch reroutes traffic to an alternate path.
   • DThe switch drops packets of that priority class.
   Reveal Solution Discussion Select an answer first
4. 28foundation · easy

   What is a potential performance concern when using overlay networks for AI workloads?

   • AOverlays are incompatible with RoCE.
   • BOverlay encapsulation adds overhead, which can reduce throughput and increase latency.
   • COverlays require a separate physical network for each tenant.
   • DOverlays cannot support multicast traffic.
   Reveal Solution Discussion Select an answer first
5. 29foundation · easy

   What characteristic of AI workload traffic makes it particularly sensitive to packet loss in the network?

   • ATraffic is primarily UDP-based and does not require reliability.
   • BTraffic is mostly small, latency-insensitive control messages.
   • CTraffic is steady and predictable, with no bursts.
   • DTraffic is bursty and uses large data transfers that can cause congestion and tail latency.
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A network architect is designing a fabric for a new AI cluster that will run both training and inference workloads. Training workloads generate synchronized, bursty east-west traffic, while inference workloads generate more predictable, steady traffic. The architect wants to ensure that the network can handle the bursty training traffic without impacting the latency-sensitive inference traffic. Which design consideration is most important?

   • AProvide sufficient buffer depth on the switches to absorb bursts
   • BDisable ECN to reduce latency
   • CUse a single spine switch to simplify the design
   • DUse a ring topology to minimize the number of hops
   Reveal Solution Discussion Select an answer first