Cisco 300-640 Practice Test 19
Question 180
The acceptance board must decide whether an AI-infrastructure response to numa locality and data-loader cpu is fit for controlled operation. The fictional operating facts are: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.
- The same-load measurement demonstrates reduced facility overhead without harming service performance.
- AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads.
- GPU duty cycle and samples-per-second improve with no host memory thrashing.
- A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
- Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
- Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
Explanation
Correct Answer
C:A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
D:A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
F:A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
Incorrect Answers
A:The same-load measurement demonstrates reduced facility overhead without harming service performance. It would not independently demonstrate that the control works: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
B:AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads. It is not the rule that supports the proposed change and acceptance test: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
E:Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior. It does not implement the governing rule for this approval: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.
Question 181
A data-center AI architecture committee considers gpu failure-domain placement in its AI Infrastructure Components and Architecture review. At the affected site: A single rack hosts all parameter workers and its PDU test interrupts every active training run. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
- AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix.
- A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency.
Explanation
Correct Answer
A:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Incorrect Answers
B:A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. It does not govern the condition described: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
C:AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix. It does not govern the condition described: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
D:A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency. It does not govern the condition described: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
Question 182
An infrastructure diagnostician is investigating gpu failure-domain placement for the AI Infrastructure Components and Architecture team. The reported incident reads: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.
- Check checkpoint frequency, write completion, replicas, integrity hashes and actual full-state restore time.
- Compare dataset manifests, content hashes, model versions, replication lag and divergent write histories.
- Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility.
- Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Explanation
Correct Answer
D:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
Incorrect Answers
A:Check checkpoint frequency, write completion, replicas, integrity hashes and actual full-state restore time. That check would not establish the decisive facts for this condition: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
B:Compare dataset manifests, content hashes, model versions, replication lag and divergent write histories. That check would not establish the decisive facts for this condition: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
C:Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility. That check would not establish the decisive facts for this condition: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Question 183
During a bounded AI-platform change-control review concerning gpu failure-domain placement, the site owner reports: A single rack hosts all parameter workers and its PDU test interrupts every active training run. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.
- Size and balance the leaf-spine fabric for the observed communication pattern before adding accelerator nodes.
- Validate and stage a supported end-to-end software bill of materials with rollback images.
- Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Explanation
Correct Answer
D:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Incorrect Answers
A:Size and balance the leaf-spine fabric for the observed communication pattern before adding accelerator nodes. That change would leave the reported problem unresolved: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
B:Validate and stage a supported end-to-end software bill of materials with rollback images. That change would leave the reported problem unresolved: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
C:Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks. That change would leave the reported problem unresolved: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Question 184
Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving gpu failure-domain placement. Their shared incident log states: A single rack hosts all parameter workers and its PDU test interrupts every active training run. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.
- Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
- Map VLAN or VRF boundaries, workload identities, firewall rules, east-west flows and attempted cross-tenant connections.
- Measure per-worker read concurrency, storage-node load, cache hit rates, queue depth and samples-per-second.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
- Reserve appropriate service classes and relieve burst contention after validating which path contributes to tail delay.
Explanation
Correct Answer
A:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
D:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
Incorrect Answers
B:Map VLAN or VRF boundaries, workload identities, firewall rules, east-west flows and attempted cross-tenant connections. It would not provide the evidence needed for the paired diagnosis and follow-up: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
C:Measure per-worker read concurrency, storage-node load, cache hit rates, queue depth and samples-per-second. It would not provide the evidence needed for the paired diagnosis and follow-up: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
E:Reserve appropriate service classes and relieve burst contention after validating which path contributes to tail delay. It is not the scoped response supported by the accompanying evidence: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
Question 185
The acceptance board must decide whether an AI-infrastructure response to gpu failure-domain placement is fit for controlled operation. The fictional operating facts are: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.
- Training storage must sustain aggregate sequential and random reads across workers, not just advertised peak bandwidth.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- Set class-based retention and quotas while keeping legally required provenance and restorable checkpoints.
- A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
- Representative small-file epochs meet the agreed open-latency and samples-per-second objectives.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Explanation
Correct Answer
B:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
D:A single rack hosts all parameter workers and its PDU test interrupts every active training run. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
F:A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
Incorrect Answers
A:Training storage must sustain aggregate sequential and random reads across workers, not just advertised peak bandwidth. It is not the rule that supports the proposed change and acceptance test: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
C:Set class-based retention and quotas while keeping legally required provenance and restorable checkpoints. It does not implement the governing rule for this approval: A single rack hosts all parameter workers and its PDU test interrupts every active training run. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
E:Representative small-file epochs meet the agreed open-latency and samples-per-second objectives. It would not independently demonstrate that the control works: A single rack hosts all parameter workers and its PDU test interrupts every active training run. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Question 186
A data-center AI architecture committee considers gpu failure-domain placement in its AI Infrastructure Components and Architecture review. At the affected site: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.
- A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads.
- High-density GPU racks must fit measured branch-circuit, PDU and failover electrical budgets under realistic power transients.
Explanation
Correct Answer
B:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Incorrect Answers
A:A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency. It does not govern the condition described: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
C:AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads. It does not govern the condition described: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
D:High-density GPU racks must fit measured branch-circuit, PDU and failover electrical budgets under realistic power transients. It does not govern the condition described: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
Question 187
An infrastructure diagnostician is investigating gpu failure-domain placement for the AI Infrastructure Components and Architecture team. The reported incident reads: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.
- Review rack nameplate ratings, measured peak draw, dual-feed loading, breaker headroom and PSU redundancy.
- Correlate request p99 with per-hop queue occupancy, drops, ECN marks and burst timing rather than average link load.
- Compare inlet temperatures, coolant flow where relevant, facility heat rejection, exhaust recirculation and throttling events.
- Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Explanation
Correct Answer
D:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
Incorrect Answers
A:Review rack nameplate ratings, measured peak draw, dual-feed loading, breaker headroom and PSU redundancy. That check would not establish the decisive facts for this condition: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
B:Correlate request p99 with per-hop queue occupancy, drops, ECN marks and burst timing rather than average link load. That check would not establish the decisive facts for this condition: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
C:Compare inlet temperatures, coolant flow where relevant, facility heat rejection, exhaust recirculation and throttling events. That check would not establish the decisive facts for this condition: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Question 188
During a bounded AI-platform change-control review concerning gpu failure-domain placement, the site owner reports: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.
- Right-size and distribute accelerator nodes across validated A/B feeds with capped power policy where required.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
- Package or shard files appropriately and scale the supported metadata path while preserving data semantics.
- Validate and stage a supported end-to-end software bill of materials with rollback images.
Explanation
Correct Answer
B:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
Incorrect Answers
A:Right-size and distribute accelerator nodes across validated A/B feeds with capped power policy where required. That change would leave the reported problem unresolved: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
C:Package or shard files appropriately and scale the supported metadata path while preserving data semantics. That change would leave the reported problem unresolved: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
D:Validate and stage a supported end-to-end software bill of materials with rollback images. That change would leave the reported problem unresolved: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Question 189
Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving gpu failure-domain placement. Their shared incident log states: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.
- Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.
- Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
- Inventory immutable datasets, intermediate artifacts, checkpoint rotations, compression and quota high-water marks.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
- Right-size and distribute accelerator nodes across validated A/B feeds with capped power policy where required.
Explanation
Correct Answer
A:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
D:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
Incorrect Answers
B:Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. It would not provide the evidence needed for the paired diagnosis and follow-up: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
C:Inventory immutable datasets, intermediate artifacts, checkpoint rotations, compression and quota high-water marks. It would not provide the evidence needed for the paired diagnosis and follow-up: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
E:Right-size and distribute accelerator nodes across validated A/B feeds with capped power policy where required. It is not the scoped response supported by the accompanying evidence: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
Question 190
The acceptance board must decide whether an AI-infrastructure response to gpu failure-domain placement is fit for controlled operation. The fictional operating facts are: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.
- A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
- The same-load measurement demonstrates reduced facility overhead without harming service performance.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
- Distributed training requires adequate effective east-west throughput between GPU workers, with oversubscription based on real collective traffic.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic.
Explanation
Correct Answer
A:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
C:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
E:A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
Incorrect Answers
B:The same-load measurement demonstrates reduced facility overhead without harming service performance. It would not independently demonstrate that the control works: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
D:Distributed training requires adequate effective east-west throughput between GPU workers, with oversubscription based on real collective traffic. It is not the rule that supports the proposed change and acceptance test: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
F:Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic. It does not implement the governing rule for this approval: A municipal analytics platform spreads GPU pods across hosts but stores all checkpoints on one unprotected local disk. Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.