Question 350
While checking a production data center change, the team finds: Distributed model training runs across GPU pods but experiences network bottlenecks because oversubscribed links carry synchronized east-west collective operations without capacity planning. Which implementation action should the engineer take first?
Explanation
Correct Answer
The root problem is that The data center fabric bandwidth and latency design do not match the high-throughput GPU collective workload. The selected change addresses it directly: Size and validate the supported high-bandwidth fabric links and traffic distribution for the AI training topology. Verify the outcome: Collective workload tests meet the required throughput and latency targets without sustained link oversubscription.
Incorrect Answers
This action concerns ai roce lossless class design, where The RDMA transport has not been mapped to an appropriate lossless traffic class with congestion safeguards. The reported failure instead stems from The data center fabric bandwidth and latency design do not match the high-throughput GPU collective workload.
This action concerns poap initial provisioning ownership, where The POAP discovery data identifies the wrong approved bootstrap artifact for the device's placement. The corrective response here is “Size and validate the supported high-bandwidth fabric links and traffic distribution for the AI training topology”; this alternative does not provide that response.
This action concerns ansible inventory variable scoping, where The playbook target inventory and variable precedence do not isolate the intended lab device group. The stem reports distributed model training runs across GPU pods but experiences network bottlenecks because oversubscribed links carry synchronized east-west collective operations without capacity planning; this control addresses a different failure condition.