Cisco 300-610 Practice Test 5
Question 40
At a two-region enterprise VXLAN fabric, the architecture standards group considers the recurring risk: A distributed job loses every epoch when one ToR switch reboots, despite redundant storage and power. The team needs a defensible design decision rather than a vendor feature checklist. Which continuing design control best prevents the same problem from recurring?
- Reassess topology and oversubscription when model parallelism or node count changes.
- Monitor utilization and facility efficiency and revisit power caps before expansion.
- Coordinate network maintenance windows with checkpoint and scheduler policies.
- Set model-version qualification tests for memory footprint and concurrency.
Explanation
Correct Answer
C:The most relevant ongoing safeguard is: Coordinate network maintenance windows with checkpoint and scheduler policies. The risk it addresses is: A distributed job loses every epoch when one ToR switch reboots, despite redundant storage and power. This check is tied to the technical principle: Training workers communicate through fabric-dependent collective operations, so network maintenance and fault containment affect effective job availability. The immediate correction remains: Design independent leaf paths and maintenance draining procedures for training traffic before selecting the topology. Periodically validate the safeguard through two different observations. Trace each worker's physical uplinks to failure domains Run an approved leaf-outage test while measuring job retry and checkpoint recovery. The continuing acceptance target is: A single planned network failure does not erase unacceptable amounts of training work.
Incorrect Answers
A:This is a continuing control for ai training east-west communication; its underlying principle is: Collective operations such as all-reduce drive high-volume east-west traffic whose latency and bandwidth matter separately from north-south application traffic. The recurring risk here is different: A distributed job loses every epoch when one ToR switch reboots, despite redundant storage and power. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: A distributed model job scales to more workers but time to train grows because inter-GPU synchronization stalls. The intended long-term acceptance condition remains: A single planned network failure does not erase unacceptable amounts of training work.
B:This is a continuing control for sustainability for gpu clusters; its underlying principle is: AI accelerators impose high electrical and thermal density; usable capacity is constrained by facility power delivery, cooling and sustainable operations. The recurring risk here is different: A distributed job loses every epoch when one ToR switch reboots, despite redundant storage and power. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: A facility specifies maximum GPU count from a vendor brochure without accounting for power and cooling headroom. The intended long-term acceptance condition remains: A single planned network failure does not erase unacceptable amounts of training work.
D:This is a continuing control for gpu memory per model replica; its underlying principle is: A model must fit within available memory per accelerator or use an approved partitioning design; aggregate cluster compute alone does not guarantee per-replica fit. The recurring risk here is different: A distributed job loses every epoch when one ToR switch reboots, despite redundant storage and power. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: An inference fleet has enough aggregate GPU FLOPS but a single model replica cannot load due to weight and context-memory pressure. The intended long-term acceptance condition remains: A single planned network failure does not erase unacceptable amounts of training work.
Question 41
At a 60-rack compute cluster, a proposed deployment has this failure mode: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The maintenance plan must not disrupt unrelated tenants. Which design correction most directly addresses this specific failure or requirement?
- Provide supported redundant subnet management with deterministic active ownership and validated failover.
- Budget receiver-facing bandwidth and buffering and validate congestion control under synchronized collectives.
- Select compatible RDMA-capable hosts, NICs and application libraries and validate the workload's use of RDMA verbs.
- Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN.
Explanation
Correct Answer
D:The relevant failure is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The technical distinction is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported. The targeted correction is: Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN. Before changing the design, check two independent observations. Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods. Accept the change only if: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
Incorrect Answers
A:This action targets infiniband subnet manager placement, where the mechanism is: InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability. That is a separate condition from the problem in this question: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. Even a successful repair of the alternative would not establish the required result: RDMA sessions across the intended routed fabric meet throughput and loss objectives. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
B:This action targets incast burst and roce microbursts, where the mechanism is: Incast creates brief concentration of traffic at an egress queue; average link utilization may hide peak buffer pressure. That is a separate condition from the problem in this question: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. Even a successful repair of the alternative would not establish the required result: RDMA sessions across the intended routed fabric meet throughput and loss objectives. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
C:This action targets rdma cpu-offload expectation, where the mechanism is: RDMA can permit NIC-assisted memory transfers with reduced CPU involvement when endpoints and middleware support the RDMA transport. That is a separate condition from the problem in this question: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. Even a successful repair of the alternative would not establish the required result: RDMA sessions across the intended routed fabric meet throughput and loss objectives. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
Question 42
At a virtualized application data center, the cause remains unconfirmed for this reported issue: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. Acceptance requires evidence from the running system, not just the saved design. Which pair of observations most directly tests the proposed diagnosis before a change?
- Inspect DSCP-to-queue mapping from endpoints to leaf; compare pause and drop counters by priority during backup bursts.
- Inspect RoCEv2 endpoint addressing and UDP flow routes; measure end-to-end loss and latency between nodes in different routed pods.
- Check installed host adapter protocol capabilities; inspect switch port family and supported link operation on each proposed path.
- Inspect ECN-marked packet and congestion-notification counters; compare no-drop queue occupancy with configured PFC triggers during controlled bursts.
Explanation
Correct Answer
B:The first observation is: Inspect RoCEv2 endpoint addressing and UDP flow routes The independent second observation is: Measure end-to-end loss and latency between nodes in different routed pods. Together they test the proposed diagnosis of this incident: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The underlying technical principle is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported. If those checks confirm it, the appropriate next action is: Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN. The expected measurable result is: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
Incorrect Answers
A:These observations investigate rdma traffic isolation; they would be useful if the symptoms were: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. That alternative rests on the principle: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. The question instead asks for evidence about: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The discriminating observations here are: Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods.
C:These observations investigate infiniband versus ethernet compatibility; they would be useful if the symptoms were: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. That alternative rests on the principle: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. The question instead asks for evidence about: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The discriminating observations here are: Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods.
D:These observations investigate roce ecn congestion signalling; they would be useful if the symptoms were: A storage-over-RDMA fabric drops traffic during moderate bursts even though PFC eventually pauses the priority. That alternative rests on the principle: ECN marks congestion before sustained queue growth and allows compatible endpoints to reduce rates; PFC handles priority-level pauses as a later protection. The question instead asks for evidence about: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The discriminating observations here are: Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods.
Question 43
At an AI-training data-center pod, a failure review must identify the central technical cause of: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. One component outage must remain inside the agreed availability budget. Which engineering principle best explains the observed behavior?
- ECN marks congestion before sustained queue growth and allows compatible endpoints to reduce rates; PFC handles priority-level pauses as a later protection.
- InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability.
- RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
- InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
Explanation
Correct Answer
C:The controlling technical principle is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported. This matters because the reported design failure is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The first diagnostic check is: Inspect RoCEv2 endpoint addressing and UDP flow routes The independent check is: Measure end-to-end loss and latency between nodes in different routed pods. These facts support the targeted intervention: Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN. Confirm that the resulting design meets this condition: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
Incorrect Answers
A:The offered principle describes roce ecn congestion signalling: ECN marks congestion before sustained queue growth and allows compatible endpoints to reduce rates; PFC handles priority-level pauses as a later protection. That is useful for a distinct incident, such as: A storage-over-RDMA fabric drops traffic during moderate bursts even though PFC eventually pauses the priority. It does not account for the reported failure: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The governing principle in this situation is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
B:The offered principle describes infiniband subnet manager placement: InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability. That is useful for a distinct incident, such as: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. It does not account for the reported failure: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The governing principle in this situation is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
D:The offered principle describes infiniband versus ethernet compatibility: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. That is useful for a distinct incident, such as: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. It does not account for the reported failure: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The governing principle in this situation is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
Question 44
At a regulated private-cloud campus, the service owner needs evidence of resolution for: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The architect must preserve the intended failure domains. Which measured result most directly proves the original requirement is satisfied?
- RDMA sessions across the intended routed fabric meet throughput and loss objectives.
- The fabric remains correctly managed after loss of one manager instance.
- RDMA receives required congestion protection without unnecessarily pausing unrelated traffic.
- Every host attaches to a mutually supported fabric with functional RDMA transport.
Explanation
Correct Answer
A:The relevant measurable success criterion is: RDMA sessions across the intended routed fabric meet throughput and loss objectives. It addresses the original reported defect: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. A saved design or configuration command alone does not establish that outcome. Observe the affected system in two independent ways: Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods. The correction being evaluated is: Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN. The underlying reason that correction matters is: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
Incorrect Answers
B:This metric would help validate a correction to infiniband subnet manager placement, rather than the original reported issue. The other issue is: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. Here the defect to be resolved is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The expected measurable outcome for the actual incident is: RDMA sessions across the intended routed fabric meet throughput and loss objectives. A positive result for the other service would not prove this requirement is met.
C:This metric would help validate a correction to rdma traffic isolation, rather than the original reported issue. The other issue is: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. Here the defect to be resolved is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The expected measurable outcome for the actual incident is: RDMA sessions across the intended routed fabric meet throughput and loss objectives. A positive result for the other service would not prove this requirement is met.
D:This metric would help validate a correction to infiniband versus ethernet compatibility, rather than the original reported issue. The other issue is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. Here the defect to be resolved is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The expected measurable outcome for the actual incident is: RDMA sessions across the intended routed fabric meet throughput and loss objectives. A positive result for the other service would not prove this requirement is met.
Question 45
At a dual-spine infrastructure project, after correction, the team wants to prevent recurrence of: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The team needs a defensible design decision rather than a vendor feature checklist. Which continuing design control best prevents the same problem from recurring?
- Revisit ECN and PFC thresholds when NIC rates or oversubscription change.
- Revalidate ECN/PFC policies with route and device changes affecting RDMA classes.
- Audit traffic classifiers and applications when new workloads join the fabric.
- Monitor subnet-manager leadership and configuration consistency.
Explanation
Correct Answer
B:The most relevant ongoing safeguard is: Revalidate ECN/PFC policies with route and device changes affecting RDMA classes. The risk it addresses is: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. This check is tied to the technical principle: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported. The immediate correction remains: Design IP reachability and consistent traffic classification for RoCEv2 instead of assuming the RDMA path must share one VLAN. Periodically validate the safeguard through two different observations. Inspect RoCEv2 endpoint addressing and UDP flow routes Measure end-to-end loss and latency between nodes in different routed pods. The continuing acceptance target is: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
Incorrect Answers
A:This is a continuing control for roce ecn congestion signalling; its underlying principle is: ECN marks congestion before sustained queue growth and allows compatible endpoints to reduce rates; PFC handles priority-level pauses as a later protection. The recurring risk here is different: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: A storage-over-RDMA fabric drops traffic during moderate bursts even though PFC eventually pauses the priority. The intended long-term acceptance condition remains: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
C:This is a continuing control for rdma traffic isolation; its underlying principle is: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. The recurring risk here is different: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. The intended long-term acceptance condition remains: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
D:This is a continuing control for infiniband subnet manager placement; its underlying principle is: InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability. The recurring risk here is different: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. The intended long-term acceptance condition remains: RDMA sessions across the intended routed fabric meet throughput and loss objectives.
Question 46
At a low-latency GPU workload fabric, a design review reports the following issue: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The maintenance plan must not disrupt unrelated tenants. Which design correction most directly addresses this specific failure or requirement?
- Budget receiver-facing bandwidth and buffering and validate congestion control under synchronized collectives.
- Classify RDMA and ordinary workloads into appropriate QoS classes with explicit bandwidth and queue policies.
- Select compatible RDMA-capable hosts, NICs and application libraries and validate the workload's use of RDMA verbs.
- Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport.
Explanation
Correct Answer
D:The relevant failure is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The technical distinction is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. The targeted correction is: Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport. Before changing the design, check two independent observations. Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path. Accept the change only if: Every host attaches to a mutually supported fabric with functional RDMA transport.
Incorrect Answers
A:This action targets incast burst and roce microbursts, where the mechanism is: Incast creates brief concentration of traffic at an egress queue; average link utilization may hide peak buffer pressure. That is a separate condition from the problem in this question: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. Even a successful repair of the alternative would not establish the required result: Every host attaches to a mutually supported fabric with functional RDMA transport. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
B:This action targets rdma traffic isolation, where the mechanism is: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. That is a separate condition from the problem in this question: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. Even a successful repair of the alternative would not establish the required result: Every host attaches to a mutually supported fabric with functional RDMA transport. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
C:This action targets rdma cpu-offload expectation, where the mechanism is: RDMA can permit NIC-assisted memory transfers with reduced CPU involvement when endpoints and middleware support the RDMA transport. That is a separate condition from the problem in this question: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. Even a successful repair of the alternative would not establish the required result: Every host attaches to a mutually supported fabric with functional RDMA transport. The applicable design change must address the diagnosed mechanism of the stated case, not the neighboring design problem.
Question 47
At a two-region enterprise VXLAN fabric, two engineers propose competing diagnoses for this condition: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. Acceptance requires evidence from the running system, not just the saved design. Which pair of observations most directly tests the proposed diagnosis before a change?
- Capture short-timescale egress queue peaks; compare incast phase timing with ECN feedback and PFC events.
- Inspect DSCP-to-queue mapping from endpoints to leaf; compare pause and drop counters by priority during backup bursts.
- Check installed host adapter protocol capabilities; inspect switch port family and supported link operation on each proposed path.
- Inspect active and standby subnet manager status; test approved manager failover and verify subnet path restoration.
Explanation
Correct Answer
C:The first observation is: Check installed host adapter protocol capabilities The independent second observation is: Inspect switch port family and supported link operation on each proposed path. Together they test the proposed diagnosis of this incident: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The underlying technical principle is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. If those checks confirm it, the appropriate next action is: Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport. The expected measurable result is: Every host attaches to a mutually supported fabric with functional RDMA transport.
Incorrect Answers
A:These observations investigate incast burst and roce microbursts; they would be useful if the symptoms were: Many GPU workers transmit synchronized results to one receiver and short-lived incast overruns receiver-side buffers. That alternative rests on the principle: Incast creates brief concentration of traffic at an egress queue; average link utilization may hide peak buffer pressure. The question instead asks for evidence about: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The discriminating observations here are: Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path.
B:These observations investigate rdma traffic isolation; they would be useful if the symptoms were: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. That alternative rests on the principle: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. The question instead asks for evidence about: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The discriminating observations here are: Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path.
D:These observations investigate infiniband subnet manager placement; they would be useful if the symptoms were: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. That alternative rests on the principle: InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability. The question instead asks for evidence about: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The discriminating observations here are: Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path.
Question 48
At a 60-rack compute cluster, the design group challenges an assumption behind this issue: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. One component outage must remain inside the agreed availability budget. Which engineering principle best explains the observed behavior?
- RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority.
- Incast creates brief concentration of traffic at an egress queue; average link utilization may hide peak buffer pressure.
- InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
- RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported.
Explanation
Correct Answer
C:The controlling technical principle is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. This matters because the reported design failure is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The first diagnostic check is: Check installed host adapter protocol capabilities The independent check is: Inspect switch port family and supported link operation on each proposed path. These facts support the targeted intervention: Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport. Confirm that the resulting design meets this condition: Every host attaches to a mutually supported fabric with functional RDMA transport.
Incorrect Answers
A:The offered principle describes rdma traffic isolation: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. That is useful for a distinct incident, such as: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. It does not account for the reported failure: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The governing principle in this situation is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
B:The offered principle describes incast burst and roce microbursts: Incast creates brief concentration of traffic at an egress queue; average link utilization may hide peak buffer pressure. That is useful for a distinct incident, such as: Many GPU workers transmit synchronized results to one receiver and short-lived incast overruns receiver-side buffers. It does not account for the reported failure: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The governing principle in this situation is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
D:The offered principle describes rocev2 routed transport design: RoCEv2 carries RDMA over UDP/IP and can be routed when end-to-end loss and congestion design are supported. That is useful for a distinct incident, such as: An AI fabric team expects RoCEv2 traffic to behave like an ordinary Layer 2-only RDMA deployment across multiple routed pods. It does not account for the reported failure: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The governing principle in this situation is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
Question 49
At a virtualized application data center, a change has been proposed for the following condition: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The architect must preserve the intended failure domains. Which measured result most directly proves the original requirement is satisfied?
- Bulk memory transfers demonstrate the intended CPU and latency benefit under load.
- Congestion is controlled with stable throughput and minimal priority pauses or loss.
- Every host attaches to a mutually supported fabric with functional RDMA transport.
- The fabric remains correctly managed after loss of one manager instance.
Explanation
Correct Answer
C:The relevant measurable success criterion is: Every host attaches to a mutually supported fabric with functional RDMA transport. It addresses the original reported defect: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. A saved design or configuration command alone does not establish that outcome. Observe the affected system in two independent ways: Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path. The correction being evaluated is: Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport. The underlying reason that correction matters is: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design.
Incorrect Answers
A:This metric would help validate a correction to rdma cpu-offload expectation, rather than the original reported issue. The other issue is: An application needs low host CPU cost for bulk transfers, but the design only adds ordinary TCP sockets on faster NICs. Here the defect to be resolved is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The expected measurable outcome for the actual incident is: Every host attaches to a mutually supported fabric with functional RDMA transport. A positive result for the other service would not prove this requirement is met.
B:This metric would help validate a correction to roce ecn congestion signalling, rather than the original reported issue. The other issue is: A storage-over-RDMA fabric drops traffic during moderate bursts even though PFC eventually pauses the priority. Here the defect to be resolved is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The expected measurable outcome for the actual incident is: Every host attaches to a mutually supported fabric with functional RDMA transport. A positive result for the other service would not prove this requirement is met.
D:This metric would help validate a correction to infiniband subnet manager placement, rather than the original reported issue. The other issue is: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. Here the defect to be resolved is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The expected measurable outcome for the actual incident is: Every host attaches to a mutually supported fabric with functional RDMA transport. A positive result for the other service would not prove this requirement is met.
Question 50
At an AI-training data-center pod, a post-incident design review focuses on preventing: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The team needs a defensible design decision rather than a vendor feature checklist. Which continuing design control best prevents the same problem from recurring?
- Audit traffic classifiers and applications when new workloads join the fabric.
- Monitor subnet-manager leadership and configuration consistency.
- Preserve adapter-to-fabric compatibility matrices for future accelerator purchases.
- Test driver, firmware and RDMA library compatibility during platform upgrades.
Explanation
Correct Answer
C:The most relevant ongoing safeguard is: Preserve adapter-to-fabric compatibility matrices for future accelerator purchases. The risk it addresses is: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. This check is tied to the technical principle: InfiniBand and Ethernet have different link-layer protocols and must use the matching network fabric or supported interworking design. The immediate correction remains: Align adapter, cable, switch and management design with the chosen InfiniBand or Ethernet RDMA transport. Periodically validate the safeguard through two different observations. Check installed host adapter protocol capabilities Inspect switch port family and supported link operation on each proposed path. The continuing acceptance target is: Every host attaches to a mutually supported fabric with functional RDMA transport.
Incorrect Answers
A:This is a continuing control for rdma traffic isolation; its underlying principle is: RDMA no-drop behavior is commonly applied only to selected traffic classes; bulk flows should not automatically share the PFC priority. The recurring risk here is different: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: A mixed-use leaf assigns bulk backup packets and latency-sensitive RoCE flows to the same queue and sees long pauses. The intended long-term acceptance condition remains: Every host attaches to a mutually supported fabric with functional RDMA transport.
B:This is a continuing control for infiniband subnet manager placement; its underlying principle is: InfiniBand fabrics use a subnet manager for discovery and fabric configuration; its placement and redundancy affect recoverability. The recurring risk here is different: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: An InfiniBand cluster has working cables and switch ports but no consistent fabric configuration after its sole subnet manager fails. The intended long-term acceptance condition remains: Every host attaches to a mutually supported fabric with functional RDMA transport.
D:This is a continuing control for rdma cpu-offload expectation; its underlying principle is: RDMA can permit NIC-assisted memory transfers with reduced CPU involvement when endpoints and middleware support the RDMA transport. The recurring risk here is different: A cluster order assumes InfiniBand host adapters can be connected directly to ordinary Ethernet switch ports without a suitable gateway. The appropriate ongoing control must prevent recurrence of that particular failure, not just the separate issue of: An application needs low host CPU cost for bulk transfers, but the design only adds ordinary TCP sockets on faster NICs. The intended long-term acceptance condition remains: Every host attaches to a mutually supported fabric with functional RDMA transport.