ECN (Explicit Congestion Notification)

PFC (Priority Flow Control)

ECN and PFC address different congestion conditions in a RoCEv2 fabric and are intended to complement rather than replace each other. ECN is the appropriate mechanism for fabric-level congestion generated by incast patterns. When several GPU or storage senders simultaneously transmit toward the same destination, switch queues begin to build. ECN-capable switches mark packets rather than immediately dropping them, allowing the receiving endpoint to generate congestion feedback so the transmitting source can reduce its sending rate.

PFC provides hop-by-hop loss prevention for the designated lossless Ethernet priority. When an endpoint or downstream interface cannot accept a burst quickly enough and a queue approaches its configured threshold, PFC transmits pause frames for that specific priority. This temporarily stops upstream transmission and prevents packet loss.

Cisco's RoCEv2 designs combine ECN and PFC precisely because ECN provides proactive end-to-end congestion management while PFC acts as the protective mechanism when queue buildup becomes severe or occurs faster than endpoint rate adaptation can respond.

WRED can provide the queue-marking mechanism associated with ECN, but the requested logical congestion-control mechanism is ECN. AFD is not the corresponding RoCEv2 lossless mechanism for either description.

Study Guide Reference: AI Infrastructure Deployment and Data Management — congestion control, RoCEv2, ECN, PFC, QoS, and lossless Ethernet.

===============

Inference is the operational phase in which a trained AI model receives new input and produces predictions, classifications, generated content, or other decisions. Cisco explicitly identifies Inference as an AI/ML workload type in DCAI objective 1.1.c. Unlike model training, which is dominated by iterative computation over large datasets, inference commonly handles production requests and is highly sensitive to response latency and throughput.

Cisco's AI POD design guidance states that inferencing workloads process live production traffic, often in real time, and evaluates them using metrics such as Time-to-First-Token, end-to-end request latency, inter-token latency, and requests per second. These characteristics directly support real-time decision making , making option C correct.

Option A concerns storage architecture rather than inference behavior. Option B incorrectly describes model architecture simplification, which inference does not inherently perform. Option D is also incorrect because inference pipelines can still require preprocessing, normalization, tokenization, or transformation of incoming data.

Study Guide Reference: 1.0 AI Fundamentals and Applications — 1.1 AI/ML workload types — 1.1.c Inference.

===============

hardware profile dlb

dlb-interface all

mode flowlet

The exhibit shows materially uneven utilization across the eight interfaces, with receive utilization ranging from approximately 32.5% to 75.3%. Cisco Dynamic Load Balancing addresses this problem by considering current path utilization rather than relying exclusively on static ECMP hashing.

The configuration begins with hardware profile dlb, which enters the DLB hardware-profile context. dlb-interface all enables DLB on all eligible Layer 3 physical interfaces. Cisco documents both an explicit interface list and the all keyword for this purpose. The document's supplied drag-and-drop answer then selects mode flowlet. Cisco defines flowlet load balancing as the default DLB mode, in which forwarding decisions use Dynamic Resilient Ethernet metrics to select less-loaded paths while maintaining packet ordering within a flowlet.

Cisco also supports per-packet DLB, which can produce even greater path utilization but can reorder packets. When NVIDIA adaptive-routing capabilities are explicitly enabled and endpoints can handle reordering, Cisco documents per-packet DLB as an optimized alternative. That additional prerequisite is not stated in this question.

Therefore, the source-derived configuration is DLB / all interfaces / flowlet .

Study Guide Reference: AI Infrastructure Deployment and Data Management — Dynamic Load Balancing, ECMP optimization, AI backend fabrics, and traffic distribution.

===============

The surveillance workload creates a substantial persistent-capacity requirement across every location. Even at the stated minimum of ten cameras, each store generates approximately 240 GB per day and roughly 3.36 TB over fourteen days before considering filesystem overhead, redundancy, additional cameras, or growth. A scalable IP-based shared-storage architecture is therefore the appropriate option.

Cisco Unified Edge is designed around integrated high-speed Ethernet networking and supports network-accessible storage through standard IP mechanisms. Cisco's validated Unified Edge guidance describes network storage using protocols such as NFS and iSCSI , while explicitly noting that Fibre Channel is not validated in that design. Unified Edge also supports software-defined and distributed storage architectures carried over its Ethernet infrastructure.

Fibre Channel would introduce specialized SAN adapters, switches, cabling, and operational complexity that are generally inappropriate for a cost-sensitive, highly distributed retail edge. InfiniBand is optimized for specialized high-performance computing rather than economical video retention. Local storage could hold surveillance data at each store, but the question explicitly requires a centralized storage solution and long-term capacity that can be managed across locations.

IP-based shared storage therefore provides the best combination of scalability, standardization, cost efficiency, and Unified Edge compatibility.

Study Guide Reference: AI Infrastructure Components and Architecture — edge AI storage, IP storage, scalability, capacity, and deployment requirements.

===============

Ethernet Network Control Policy

Ethernet Network Group Policy

Ethernet QoS Policy

Ethernet Adapter Policy

The mappings correspond to the functional boundaries of Cisco Intersight vNIC policies. Ethernet Network Control Policy manages link-layer behavior including CDP and LLDP. Cisco's Intersight documentation explicitly identifies CDP, LLDP transmit/receive, MAC registration, and uplink-failure behavior as Network Control settings.

Ethernet Network Group Policy controls VLAN membership for FI-attached servers, including the set of allowed VLANs and the optional native VLAN. Ethernet QoS Policy supplies performance-oriented vNIC characteristics such as MTU, rate limiting, traffic priority, CoS, and burst parameters. Cisco's current RoCE configuration guide lists MTU and rate-limit values directly under Ethernet QoS configuration.

Finally, Ethernet Adapter Policy contains adapter-level behavior and performance parameters, including RoCE configuration used for high-performance RDMA interfaces.

Ethernet Network Policy is the unused choice. Cisco documents that policy primarily for VLAN configuration on standalone servers, whereas Ethernet Network Group Policy is the relevant VLAN policy for Fabric Interconnect-attached UCS X-Series vNICs.

Study Guide Reference: AI Infrastructure Deployment and Data Management — UCS vNIC templates, Ethernet policies, QoS, VLAN configuration, and RoCE.

===============

In Cisco Intersight Managed Mode, Fabric Interconnect physical port roles and port-channel definitions are configured through the Port Policy . Cisco's current Intersight configuration documentation states that a Port Policy configures ports and port roles on a Fabric Interconnect and permits administrators to enable or disable individual ports or port channels. Supported roles include Ethernet uplinks and Ethernet uplink port channels.

The Port Policy is subsequently associated with the UCS Domain Profile. Cisco documents that a Domain Profile configures an FI pair through reusable policies and controls ports, port channels, VLANs, and VSANs. During Domain Profile creation, the administrator attaches the appropriate Port Policy to each Fabric Interconnect.

A LAN Connectivity Policy is used primarily at the server-profile level to define vNIC connectivity. A Network Connectivity Policy configures DNS-related server-management settings rather than FI uplink topology. An Ethernet Network Group Policy defines VLAN groups associated with Ethernet connectivity and can be referenced by uplink configurations, but it does not itself create the FI port channel.

Consequently, the policy that must be changed to build the multiple-100-GbE uplink port channel is the Port Policy .

Study Guide Reference: AI Infrastructure Deployment and Data Management — Intersight Managed Mode, UCS Domain Profiles, Fabric Interconnect port configuration, and uplink resiliency.

===============

The uploaded source marks C and D . GPU Memory Utilization in option D is clearly relevant: GPU memory pressure is a critical accelerator metric because model weights, activations, KV caches, tensors, and training state consume GPU-attached memory. Cisco AI monitoring material explicitly includes GPU memory utilization among the telemetry used for AI infrastructure visibility.

Option C, Active CPU Utilization , is useful as a supporting performance metric because AI workloads can become CPU-bound during data preprocessing, tokenization, orchestration, input-pipeline execution, or host-to-device preparation. A CPU bottleneck can therefore leave GPUs underutilized even when sufficient accelerator capacity exists.

However, the question as written has a notable construction issue: Cisco documentation treats GPU utilization itself as a principal metric for evaluating accelerator utilization, but that metric is absent from the available answer choices. Cisco current material explicitly identifies GPU utilization and GPU memory utilization as AI monitoring measurements.

Therefore, CD is retained because it is the answer supplied by the source , but the technically ideal metric pair would be GPU Utilization + GPU Memory Utilization if GPU Utilization were offered.

Study Guide Reference: AI Infrastructure Operations and Troubleshooting — GPU utilization, GPU memory utilization, host-resource bottlenecks, and AI performance monitoring.

===============

Cisco Nexus Hyperfabric's full-stack AI infrastructure integrates compute, GPU acceleration, high-performance networking, optional enterprise storage, and centralized cloud-based lifecycle management. Cisco identifies the principal components as the Nexus Hyperfabric cloud controller, Cisco 6000/Nexus switching platforms, Cisco UCS GPU/DPU servers equipped with NVIDIA accelerators, and optional Cisco UCS-based VAST storage.

Option C therefore contains the essential architecture: Cisco UCS servers with NVIDIA GPUs, Cisco switching, optional VAST Data storage, and cloud-managed networking .

Option A omits the cloud-controller architecture and does not specifically identify the NVIDIA-based accelerated compute model that characterizes the full-stack AI solution. Option B incorrectly describes cloud management as optional. Cloud-based fabric design, deployment, visibility, and lifecycle control are fundamental elements of Nexus Hyperfabric. Option D omits GPU compute entirely and reduces the solution to networking and software-defined storage, which does not represent the integrated AI stack.

Cisco's architecture is intended to reduce the complexity associated with constructing AI infrastructure manually. Validated cluster templates coordinate compute, networking, storage, optics, cabling, and lifecycle operations while aligning with NVIDIA enterprise AI reference architectures.

Study Guide Reference: AI Infrastructure Components and Architecture — Cisco Hyperfabric, GPU compute, AI networking, storage integration, scalability, and cloud management.

===============