ExamworthyExamworthy
Open the appStart free

NCA-AIIO - AI Infrastructure - Section 2.7

Determine networking requirements for AI workloads.

Determine the bandwidth, latency, and topology requirements of distributed AI training workloads, where all-reduce collective operations place heavy demand on the compute fabric. Distinguish between the compute network, used for GPU-to-GPU communication, and the storage and management networks, and recognise when RDMA over Converged Ethernet (RoCE) or InfiniBand is appropriate.

Practice question for this objective

Free sampleAI Infrastructuremedium

A data centre operations team notices that NFS checkpoint writes from a 32-node training job are causing packet drops on the same fabric used for gradient all-reduce traffic. Which network design change best addresses this interference?

  • AIncrease the MTU on the shared fabric to reduce packet overhead for both workload types simultaneously
  • BEnable flow control (PFC) on the shared fabric so checkpoint traffic yields to gradient all-reduce flows
  • CReplace InfiniBand with 400 GbE Ethernet so the single shared fabric has enough capacity for both traffic types
  • DProvision a dedicated high-bandwidth storage network for checkpoint and dataset I/O, separate from the GPU compute fabric Correct
Recognise that separating the compute fabric from storage and management networks prevents I/O contention from degrading distributed training performance. AI training clusters generate two distinct types of high-volume network traffic: east-west gradient all-reduce traffic between GPUs during training, and north-south storage traffic for reading datasets and writing checkpoints. When these flows share a fabric, storage bursts compete with collective operations for switch buffers and link bandwidth, introducing latency and packet drops that slow gradient synchronisation. The standard design response is to provision physically or logically separate networks - a dedicated high-bandwidth, low-latency compute fabric for GPU-to-GPU traffic, and a separate storage network for persistent I/O - so each plane is independently optimised and neither interferes with the other.

Why A is wrong: Increasing MTU (jumbo frames) can improve throughput for bulk transfers but does not prevent checkpoint I/O and gradient traffic from competing for the same switch buffers. When two high-bandwidth flows share a fabric, they still contend for bandwidth regardless of frame size, so the interference persists.

Why B is wrong: Priority-based flow control can prioritise gradient traffic over storage traffic on a shared link, but it does not eliminate contention for total fabric bandwidth. When the aggregate demand from both flows exceeds link capacity, PFC will pause the lower-priority flow, increasing checkpoint latency and risking deadlock if misconfigured. Separation is more reliable than priority alone.

Why C is wrong: Upgrading to a faster shared medium may reduce congestion temporarily but reintroduces the same competition at higher line rates as the cluster or checkpoint frequency grows. A single shared fabric is still a single point of contention; the architectural solution is separation of traffic planes, not simply provisioning more bandwidth on a shared medium.

Why D is correct: Separating the storage network from the compute fabric ensures that bulk checkpoint writes cannot consume bandwidth or inject congestion into the fabric that gradient all-reduce depends on. Each network can then be independently sized and tuned for its workload: low-latency, non-blocking for compute and high-throughput for storage, eliminating cross-workload interference entirely.

See more NCA-AIIO practice questions, answers explained.

More in this domain

Back to all AI Infrastructure objectives, or the NCA-AIIO cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.