3.2 AI Network Design Use Cases (such as Machine Learning, Large Language Models, and Pattern Recognition) 400-007 Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   Which network characteristic is most critical for a pattern recognition system that performs real-time analytics on streaming sensor data?

   • ALow bandwidth to reduce cost
   • BLow and predictable latency
   • CHigh jitter to avoid congestion
   • DMaximum throughput regardless of latency
   Reveal Solution Discussion Select an answer first
2. 12foundation · easy

   Which statement accurately describes a network requirement commonly associated with AI inference workloads?

   • AInference workloads are insensitive to network latency because models are pre-trained.
   • BInference traffic is typically north-south and can tolerate high jitter.
   • CInference typically requires low latency to return results quickly to users or downstream systems.
   • DInference requires massive sustained bandwidth between GPUs for gradient exchange.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A healthcare organization is building an LLM training cluster to process sensitive patient data. The security team requires that the training data and model checkpoints be isolated from the rest of the corporate network. The data science team needs to access the cluster for development and monitoring. Which network design approach best meets both the security and accessibility requirements?

   • AConnect the training cluster directly to the corporate core switch and rely on host-based firewalls on each GPU server
   • BUse a VRF to separate the training cluster traffic and allow all traffic from the corporate network to reach it
   • CDeploy the training cluster in a public cloud and use a VPN for data science team access
   • DPlace the training cluster in a dedicated VLAN with a stateful firewall that allows only SSH from the data science team's subnet
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   Which telemetry data is most useful for detecting network congestion during AI training?

   • AThe number of active users on the corporate Wi-Fi
   • BThe ambient temperature of the data center
   • CCPU utilization of the management servers
   • DQueue depth and packet drop counters on switch interfaces
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A large internet company is designing a network for a new LLM training cluster. The cluster uses a mix of long-lived flows for model synchronization and short-lived flows for data loading. The network team must ensure that the long-lived flows do not starve the short-lived flows. The fabric uses a spine-and-leaf topology with standard buffers. Which congestion control and flow scheduling approach is most effective?

   • AEnable explicit congestion notification (ECN) and use a data center TCP variant like DCTCP to handle congestion
   • BUse priority flow control (PFC) to prevent buffer overflow and rely on TCP timeouts to recover from congestion
   • CIncrease the buffer size on all switches to absorb bursts and reduce the impact of long-lived flows
   • DImplement a centralized flow scheduler that prioritizes short-lived flows and uses weighted fair queuing (WFQ) at the switches
   Reveal Solution Discussion Select an answer first