I85PREP
⌘ K
Cisco ENCC (300-440) study materials
Cisco ENCC (300-440) study resources

Cisco ENCC (300-440) practice questions

One question at a time. Build your Cisco ENCC (300-440) understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Infrastructure

In a Cisco AI infrastructure fabric, which component is primarily responsible for providing lossless transport for RDMA traffic between GPU servers?

  1. A priority flow control enabled Ethernet fabric with PFC and ECN Correct answer.
  2. A standard spanning-tree topology with no congestion control
  3. A storage area network using Fibre Channel zoning
  4. A routed WAN edge using BGP only

Explanation

RoCEv2 requires a lossless or near-lossless Ethernet underlay so that RDMA packets are not dropped during congestion. Priority Flow Control (PFC) pauses specific traffic classes, and Explicit Congestion Notification (ECN) marks packets so receivers can signal congestion back to senders. Together they form the standard congestion management approach for AI/GPU clusters.

Question 2AI Infrastructure

Which statement best describes the role of ECN in a RoCEv2 fabric?

  1. It encrypts RDMA payloads between GPU servers
  2. It load-balances flows across equal-cost paths using a hash
  3. It marks packets at congested queues so receivers can send congestion notifications to senders Correct answer.
  4. It assigns VLAN tags to storage traffic

Explanation

Explicit Congestion Notification works by having a congested switch queue mark the ECN bits in the IP header instead of dropping the packet. The receiver echoes this marking back to the sender, which then reduces its transmission rate. This avoids the retransmission cost that packet drops would impose on RDMA flows.

Question 3AI Infrastructure

A network engineer is designing a spine-leaf fabric for GPU training nodes and wants to minimize oversubscription at the leaf. Which design choice best supports that goal?

  1. Aggregating all leaf uplinks into a single 10G link
  2. Providing non-blocking uplink capacity from each leaf equal to its downlink capacity Correct answer.
  3. Relying on spanning tree to block redundant paths
  4. Using a single spine with no redundancy

Explanation

AI training traffic is often bursty and many-to-many, so a leaf with fewer uplink than downlink bandwidth becomes a bottleneck. A non-blocking design sizes uplinks to match downlinks, keeping oversubscription at 1:1. This preserves throughput for collective operations such as all-reduce.

Question 4AI Infrastructure

Which protocol is commonly used to advertise tenant or overlay routes in a Cisco EVPN-VXLAN data center fabric?

  1. IS-IS with no extensions
  2. BGP with EVPN address family Correct answer.
  3. RIPv2
  4. OSPFv2 only

Explanation

EVPN uses BGP as its control plane, carrying MAC and IP reachability information in the EVPN address family. This enables efficient multihoming, mobility, and suppression of unnecessary flooding. Plain OSPF, RIP, or IS-IS without extensions cannot carry EVPN route types.

Question 5AI Infrastructure

In a VXLAN fabric, what is the function of the VXLAN Network Identifier (VNI)?

  1. It sets the MTU for the underlay
  2. It identifies the overlay segment, analogous to a VLAN ID but with a much larger space Correct answer.
  3. It determines the BGP autonomous system number
  4. It identifies the physical switch port used for uplinks

Explanation

The VNI is a 24-bit identifier that maps traffic to a specific overlay network segment. It allows far more segments than the 12-bit VLAN ID space and is carried in the VXLAN header. VTEPs use the VNI to decide which overlay segment a frame belongs to.

Question 6AI Infrastructure

A GPU cluster experiences intermittent RDMA retransmissions during large collective operations. Which two telemetry signals would most directly help isolate the cause? (Choose the best single answer.)

  1. NTP offset and syslog severity levels
  2. CDP neighbor lists and VTP domain names
  3. Interface PFC pause counters and ECN marked-packet counters Correct answer.
  4. DNS query logs and DHCP lease times

Explanation

PFC pause counters reveal whether a queue is being paused frequently, and ECN marked-packet counters show congestion signaling along the path. Correlating these with the timing of retransmissions points to the congested queue or link. DNS, DHCP, NTP, CDP, and VTP data do not describe congestion behavior.

Question 7AI Infrastructure

Which feature allows a Cisco Nexus switch to buffer and absorb bursts of traffic so that PFC pauses are less frequently triggered?

  1. Enabling storm control on server ports
  2. A smaller per-port buffer with static limits
  3. Disabling ECN on all queues
  4. A deep shared buffer with dynamic thresholding Correct answer.

Explanation

Dynamic buffer thresholding lets a switch allocate shared buffer space to congested queues on demand. This absorbs microbursts before they trigger PFC pauses, improving RDMA performance. Static small buffers and disabled ECN tend to increase drops or pauses.

Question 8AI Infrastructure

What is the main purpose of a Cisco Nexus Dashboard in an AI infrastructure deployment?

  1. To provide centralized management, automation, and telemetry across data center fabrics Correct answer.
  2. To act as a DNS authoritative server for the enterprise
  3. To terminate VPN tunnels for remote branch users
  4. To provide wireless LAN controller functions

Explanation

Nexus Dashboard is a management platform that centralizes fabric configuration, automation, and telemetry collection for data center networks. It gives operators a single view of multiple fabrics and sites. VPN termination, DNS, and wireless control are handled by other product families.

Question 9AI Infrastructure

Which Cisco technology provides a policy-driven, intent-based approach to automating data center fabric configuration?

  1. Cisco IOS classic static routes
  2. Cisco Umbrella DNS filtering
  3. Cisco Nexus Dashboard Fabric Controller with intent-based policies Correct answer.
  4. Cisco AnyConnect posture assessment

Explanation

Nexus Dashboard Fabric Controller (formerly DCNM) provides automated fabric lifecycle management driven by declared intent and policy. It handles underlay and overlay provisioning across the fabric. The other options address routing, endpoint posture, or DNS security rather than fabric automation.

Question 10AI Infrastructure

Which consideration is most important when sizing the underlay MTU for a VXLAN fabric carrying RDMA traffic?

  1. The underlay MTU must be smaller than the server MTU
  2. The underlay must accommodate the VXLAN encapsulation overhead plus the original frame size Correct answer.
  3. The underlay MTU must be set to 576 bytes
  4. The underlay MTU must equal the VNI value

Explanation

VXLAN adds outer headers, so the underlay must carry the original frame plus overhead without fragmentation. If the underlay MTU is too small, encapsulated packets are dropped or fragmented, harming performance. The VNI value and a tiny MTU have no bearing on correct sizing.

Question 11AI Infrastructure

A network team wants to detect microbursts that last only microseconds on 400G links serving GPU nodes. Which telemetry approach is most appropriate?

  1. Hardware-based streaming telemetry with high-resolution interface counters Correct answer.
  2. Reviewing monthly syslog summaries
  3. Polling interface counters every five minutes with SNMP
  4. Collecting NetFlow records with a one-hour export interval

Explanation

Microbursts occur on very short timescales, so counters must be sampled at high resolution in hardware to observe them. Streaming telemetry pushes fine-grained data continuously to collectors. Five-minute SNMP polling, monthly logs, and hourly flow exports average away the bursts.

Question 12AI Infrastructure

Which statement about RoCEv2 is accurate?

  1. It encapsulates RDMA over UDP/IP and can be routed across Layer 3 networks Correct answer.
  2. It requires Fibre Channel as the transport
  3. It operates only within a single Layer 2 broadcast domain
  4. It encrypts all traffic by default using IPsec

Explanation

RoCEv2 places RDMA payloads inside UDP/IP, which allows the traffic to be routed across Layer 3 fabrics. This is a key advantage over RoCEv1, which was limited to a single Layer 2 domain. RoCEv2 does not inherently provide encryption.

Question 13AI Infrastructure

In a Cisco AI fabric, what does a PFC watchdog feature do?

  1. It measures the round-trip time of BGP updates
  2. It synchronizes clocks across the fabric
  3. It enforces QoS marking on storage traffic
  4. It detects and recovers from stuck pause frames that would otherwise halt a queue indefinitely Correct answer.

Explanation

A stuck PFC pause can block a queue for a long period and stall the whole traffic class. The PFC watchdog monitors pause duration and takes recovery action such as dropping or resetting the affected queue. This prevents a single misbehaving link from halting RDMA traffic.

Question 14AI Infrastructure

Which Cisco solution is designed to provide visibility into application and flow behavior across a data center fabric without deploying physical taps on every link?

  1. Cisco Nexus Dashboard Insights with flow telemetry Correct answer.
  2. Cisco SD-WAN vManage templates
  3. Cisco ISE posture policies
  4. Cisco Prime Infrastructure wireless maps

Explanation

Nexus Dashboard Insights collects flow and telemetry data from fabric switches to provide application visibility and anomaly detection. It avoids the cost and complexity of physical taps on each link. ISE, SD-WAN, and Prime Infrastructure address different domains.

Question 15AI Infrastructure

An engineer observes that a single elephant flow is saturating one member of a 4-way ECMP bundle while other members are underused. Which mechanism most directly addresses this imbalance?

  1. Increasing the spanning-tree priority of the switch
  2. Reducing the MTU on the congested link
  3. Flow-based load balancing with a better hash or per-packet spraying such as dynamic load balancing Correct answer.
  4. Disabling ECN on the congested link

Explanation

ECMP hashing can map a large flow to one member, causing imbalance. Improved hashing or per-packet spraying spreads traffic more evenly across members. Spanning-tree priority, ECN, and MTU changes do not resolve flow-to-link imbalance.

Question 16AI Infrastructure

Which protocol is typically used to distribute precise time to switches and servers so that telemetry timestamps align across an AI cluster?

  1. PTP (IEEE 1588) Correct answer.
  2. LACP
  3. RIPv2
  4. VTP

Explanation

Precision Time Protocol provides sub-microsecond clock synchronization suitable for correlating telemetry across many devices. Accurate timestamps are essential when analyzing microbursts and congestion events. RIP, VTP, and LACP serve routing, VLAN management, and link aggregation respectively.

Question 17AI Infrastructure

A fabric is being prepared for a large GPU training cluster. Which QoS design element is most important for protecting RDMA traffic during congestion?

  1. Marking RDMA traffic as scavenger class
  2. Placing RDMA traffic in the default best-effort class
  3. Disabling QoS classification on all server ports
  4. Assigning RDMA traffic to a dedicated no-drop class with PFC enabled Correct answer.

Explanation

RDMA traffic should be placed in a dedicated no-drop class so PFC can pause only that class without affecting others. This isolates lossless behavior and protects latency-sensitive GPU traffic. Best-effort, scavenger, or disabled classification would expose RDMA to drops.

Question 18AI Infrastructure

Which statement about the Cisco Nexus 9000 series in an AI data center is accurate?

  1. It supports high-density 400G and 800G ports suited to GPU server connectivity Correct answer.
  2. It is limited to 1G access ports only
  3. It runs only Fibre Channel protocols
  4. It cannot participate in VXLAN fabrics

Explanation

Nexus 9000 switches include high-bandwidth platforms with 400G and 800G interfaces that match the throughput needs of GPU servers. They support VXLAN, EVPN, and advanced congestion management. The other statements contradict the platform's capabilities.

Question 19AI Infrastructure

What is the primary benefit of using a two-tier spine-leaf topology over a traditional three-tier core-aggregation-access design for AI fabrics?

  1. It removes the need for redundant power
  2. It reduces the number of physical links required to one per server
  3. It provides predictable, low-latency any-to-any paths with equal-cost multipathing Correct answer.
  4. It eliminates the need for routing protocols

Explanation

A spine-leaf design gives every leaf an equal-cost path to every other leaf, so server-to-server latency is consistent. This predictability suits the many-to-many communication patterns of distributed training. It does not reduce links to one per server or remove routing or power redundancy needs.

Question 20AI Infrastructure

Which mechanism helps prevent a single misbehaving server from consuming all shared buffer resources on a switch port?

  1. Removing PFC from the port
  2. Disabling all QoS on the port
  3. Increasing the MTU on the port
  4. Per-port and per-queue buffer limits with dynamic thresholds Correct answer.

Explanation

Buffer limits constrain how much shared buffer a port or queue can occupy, preventing one flow from starving others. Dynamic thresholds adjust these limits based on current utilization. Disabling QoS, raising MTU, or removing PFC does not provide fair buffer allocation.

Question 21AI Infrastructure

An engineer needs to verify that ECN is functioning end to end on a RoCEv2 path. Which observation best confirms correct operation?

  1. All packets are dropped at the congested queue
  2. Congested queues mark packets with CE and senders reduce rate in response to congestion notifications Correct answer.
  3. PFC pause frames are never generated under any load
  4. The receiver increases its window size when marks arrive

Explanation

Correct ECN operation means congested queues set the Congestion Experienced codepoint rather than dropping, and receivers return congestion notifications so senders slow down. Seeing rate reduction in response to marks confirms the loop works. Dropped packets or ignored marks indicate misconfiguration.

Question 22AI Infrastructure

Which capability of Cisco Nexus Dashboard Fabric Controller supports consistent configuration of many leaf switches?

  1. Wireless controller profile push
  2. Manual CLI entry on each switch individually
  3. Per-switch static route configuration
  4. Template-based fabric provisioning with policy compliance Correct answer.

Explanation

Template-based provisioning lets an operator define configuration once and apply it consistently across many switches. Policy compliance then detects drift from the intended state. Manual per-switch CLI, static routes, and wireless profiles do not provide scalable fabric consistency.

Question 23AI Infrastructure

Which factor most directly determines the achievable bisection bandwidth of an AI spine-leaf fabric?

  1. The number of VLANs configured on the leaves
  2. The number and speed of spine-to-leaf links relative to the number of server-facing ports Correct answer.
  3. The syslog severity level configured on the spines
  4. The DNS domain name used by the servers

Explanation

Bisection bandwidth is the capacity available across the middle of the fabric, which depends on the spine-to-leaf link count and speed compared with the total server-facing capacity. More or faster uplinks raise bisection bandwidth. VLAN counts, syslog levels, and DNS names do not affect capacity.

Question 24AI Infrastructure

Which Cisco feature allows an operator to apply consistent security policy to traffic between overlay segments in a VXLAN fabric?

  1. Distributed policy enforcement at the leaf VTEP using a security policy model Correct answer.
  2. Disabling inter-VNI routing entirely
  3. Using spanning tree to isolate segments
  4. Relying on physical firewall appliances for all east-west traffic

Explanation

Policy enforcement at the leaf VTEP lets security rules be applied close to the endpoints for east-west traffic. This avoids tromboning all inter-segment flows through a central appliance. Spanning tree and disabled routing do not provide policy control.

Question 25AI Infrastructure

Which statement best explains why lossless Ethernet for RDMA requires careful coordination between PFC and ECN?

  1. PFC prevents drops by pausing senders, while ECN reduces pause propagation by signaling senders to slow down earlier Correct answer.
  2. PFC and ECN are mutually exclusive and must never be enabled together
  3. ECN pauses the link while PFC marks packets
  4. Both mechanisms only affect broadcast traffic

Explanation

PFC provides a hard backpressure mechanism that stops traffic before buffers overflow, but excessive pausing can spread congestion through the fabric. ECN gives earlier, gentler feedback so senders reduce rate before pauses are needed. Using both together balances lossless delivery with congestion control.

Question 26AI Infrastructure

In a Cisco Nexus 9000-based AI fabric, which interface type is used to connect GPU servers to leaf switches while carrying RoCEv2 traffic?

  1. A routed Layer 3 port configured as a point-to-point link Correct answer.
  2. A trunk port carrying all VLANs in the fabric
  3. A Layer 2 access port in the default VLAN
  4. A promiscuous span destination port

Explanation

Modern AI fabrics typically use routed point-to-point links between servers and leaves so that ECMP and per-flow hashing operate at Layer 3 and broadcast domains stay small. Trunk and access designs introduce Layer 2 flooding and STP concerns that complicate lossless behavior.

Question 27AI Infrastructure

Which Cisco Nexus 9000 capability allows a switch to absorb short traffic bursts without dropping packets so that PFC pauses are triggered less often?

  1. Egress shaping
  2. Shared buffer with dynamic thresholding Correct answer.
  3. Storm control
  4. Ingress rate limiting

Explanation

Dynamic buffer thresholding lets a port claim additional shared buffer when a burst arrives, absorbing the burst before the queue overflows and PFC is needed. Rate limiting, shaping, and storm control all restrict or police traffic rather than absorb bursts.

Question 28AI Infrastructure

Which protocol does a Cisco Nexus switch use to advertise overlay endpoint reachability in a VXLAN EVPN fabric?

  1. PIM-SM
  2. OSPFv2
  3. IS-IS
  4. MP-BGP with the EVPN address family Correct answer.

Explanation

EVPN uses MP-BGP to distribute MAC and IP reachability information for overlay endpoints, including the type-2 and type-5 routes used for hosts. OSPF, IS-IS, and PIM handle underlay routing or multicast and do not carry EVPN endpoint advertisements.

Question 29AI Infrastructure

What is the primary role of the Cisco Nexus Dashboard in an AI infrastructure deployment?

  1. Terminating RoCEv2 traffic from GPU servers
  2. Replacing the underlay routing protocol
  3. Acting as the default gateway for all server VLANs
  4. Providing a unified operations console for fabric controllers and insights services Correct answer.

Explanation

Nexus Dashboard aggregates fabric controllers, assurance, and insights services into a single management plane for operations teams. It does not forward data-plane traffic, act as a gateway, or run the underlay routing.

Question 30AI Infrastructure

Which statement describes how RoCEv2 differs from RoCEv1?

  1. RoCEv2 uses TCP to guarantee delivery of RDMA payloads
  2. RoCEv2 encapsulates RDMA over UDP/IP and can be routed across Layer 3 Correct answer.
  3. RoCEv2 requires a lossless Layer 2 domain and cannot be routed
  4. RoCEv2 removes the need for ECN and PFC entirely

Explanation

RoCEv2 places the RDMA payload in UDP with an IP header, which makes it routable across a Layer 3 fabric and enables ECMP. RoCEv1 was tied to Ethernet Layer 2 and could not be routed, and RoCEv2 still relies on ECN and PFC for lossless behavior.

Question 31AI Infrastructure

Which Cisco feature detects and mitigates PFC deadlocks by temporarily disabling a paused queue when a pause persists too long?

  1. UDLD
  2. BPDU guard
  3. Storm control
  4. PFC watchdog Correct answer.

Explanation

PFC watchdog monitors how long a queue remains paused and, if the pause exceeds a configured threshold, drops the queue's traffic or resets the pause to break the deadlock. Storm control, UDLD, and BPDU guard address unrelated Layer 2 protection functions.

Question 32AI Infrastructure

Which QoS marking should an operator apply to RoCEv2 traffic so that it can be placed in a lossless, no-drop queue on Cisco Nexus switches?

  1. DSCP 0 mapped to the default class
  2. DSCP EF (46) mapped to a no-drop class Correct answer.
  3. DSCP CS5 (40) mapped to a strict-priority class
  4. DSCP CS1 (8) mapped to a scavenger class

Explanation

RoCEv2 is typically marked with a high-priority DSCP such as EF and mapped to a no-drop queue with PFC enabled. Best-effort, scavenger, and strict-priority classes do not provide the lossless treatment RDMA requires.

Question 33AI Infrastructure

Which statement describes the role of the Cisco Nexus Dashboard Fabric Controller in an AI fabric?

  1. It provides DNS resolution for compute nodes
  2. It automates and manages fabric lifecycle and configuration across many switches Correct answer.
  3. It performs line-rate forwarding of GPU traffic
  4. It replaces the need for an underlay routing protocol

Explanation

Nexus Dashboard Fabric Controller is the management and automation layer that provisions and maintains fabric switches at scale. It does not forward data-plane traffic, provide DNS, or eliminate underlay routing.

Question 34AI Infrastructure

An engineer wants to confirm that ECN is marking rather than dropping packets on a RoCEv2 path. Which counter observation best supports that conclusion?

  1. PFC pause counters increase steadily on the egress port
  2. Input errors increase on the receiving interface
  3. Output queue drops increase while ECN-marked counters stay at zero
  4. ECN-marked packet counters increase while output queue drops stay near zero Correct answer.

Explanation

When ECN is working, congestion is signaled by setting the ECN bits in the IP header, so marked counters rise while tail drops remain low. Rising drops with no markings indicate ECN is not engaging, and interface errors or rising PFC pauses indicate different problems.

Question 35AI Infrastructure

An engineer observes that a single large flow is saturating one member of a 4-way ECMP bundle while other members are underused. Which approach best improves flow distribution for RoCEv2 traffic?

  1. Lower the MTU on all member links
  2. Increase the PFC pause threshold
  3. Enable flowlet-based or adaptive load balancing on the ECMP bundle Correct answer.
  4. Disable ECMP and use a single high-capacity link

Explanation

Flowlet or adaptive load balancing rehashes flows when gaps appear between bursts, which spreads long-lived flows across more ECMP members. Removing ECMP or adjusting PFC and MTU does not change how flows are distributed.

Question 36AI Infrastructure

A design team is sizing the underlay MTU for a VXLAN fabric that will carry RoCEv2. What must be accounted for when setting the underlay MTU?

  1. No additional overhead because VXLAN is transparent
  2. The VXLAN, UDP, and IP outer headers added to the original frame Correct answer.
  3. Only the RoCEv2 UDP header
  4. Only the Ethernet preamble and FCS

Explanation

VXLAN encapsulation adds outer Ethernet, IP, UDP, and VXLAN headers, so the underlay MTU must be large enough to carry the original frame plus that overhead without fragmentation. Considering only one header or assuming transparency leads to drops on encapsulated packets.

Question 37AI Infrastructure

Which Cisco Nexus Dashboard Insights capability is most useful when investigating microbursts on 400G links serving GPU nodes?

  1. Role-based access control auditing
  2. Firmware upgrade orchestration
  3. Static configuration compliance checks
  4. Flow and telemetry analytics with sub-second granularity Correct answer.

Explanation

Microbursts occur in very short intervals, so telemetry and flow analytics with fine-grained sampling are needed to detect them. Compliance, upgrade, and access-control features operate on management data and cannot reveal transient congestion.

Question 38AI Infrastructure

A network team wants to detect microbursts lasting only microseconds on 400G links serving GPU nodes. Which telemetry approach is best suited?

  1. SNMP polling every five minutes
  2. High-frequency streaming telemetry with hardware timestamping Correct answer.
  3. Syslog messages generated on interface resets
  4. NetFlow export with one-minute aggregation

Explanation

Microsecond bursts require high-frequency streaming telemetry backed by hardware timestamps so samples align across devices. Five-minute polling, event-driven syslog, and coarse flow aggregation all miss short-lived spikes.

Question 39AI Infrastructure

Which mechanism helps prevent one misbehaving server from consuming all shared buffer resources on a switch port?

  1. Disabling PFC on all ports
  2. Enabling jumbo frames on the uplink
  3. Per-port and per-queue buffer limits with dynamic thresholds Correct answer.
  4. Increasing the global shared buffer without limits

Explanation

Per-port and per-queue limits cap how much of the shared buffer a single port or queue can claim, protecting other traffic from a misbehaving sender. Unlimited shared buffering, disabling PFC, or changing MTU does not isolate buffer consumption.

Question 40AI Infrastructure

Which approach provides consistent security policy between overlay segments in a VXLAN EVPN fabric?

  1. Disabling routing between VRFs
  2. Distributed anycast gateways with policy enforcement on leaf switches Correct answer.
  3. Using static MAC ACLs on spine switches
  4. Relying on VLAN isolation alone

Explanation

Placing anycast gateways on the leaves allows policy to be enforced close to the workloads as traffic crosses between VRFs or segments. VLAN isolation, spine ACLs, and disabled routing do not deliver scalable, consistent inter-segment policy.

Question 41AI Infrastructure

Which protocol is commonly used to distribute precise time to switches and servers so that telemetry timestamps align across an AI fabric?

  1. SNTP
  2. NTP
  3. PTP Correct answer.
  4. RADIUS

Explanation

Precision Time Protocol provides sub-microsecond synchronization suitable for correlating telemetry across many devices. NTP and SNTP provide coarser accuracy, and RADIUS is an authentication protocol.

Question 42AI Infrastructure

Which statement about Cisco Nexus 9000 series switches in an AI data center is accurate?

  1. They support high-density 400G and 800G ports suitable for GPU server connectivity Correct answer.
  2. They are limited to 10G and 25G port speeds
  3. They cannot run VXLAN EVPN
  4. They do not support PFC

Explanation

Nexus 9000 platforms include models with high-density 400G and 800G interfaces designed for AI and high-performance workloads. They support VXLAN EVPN and PFC, and they are not limited to 10G or 25G speeds.

Question 43AI Infrastructure

An engineer needs to verify that PFC is functioning on a leaf port carrying RoCEv2 traffic. Which observation best confirms correct operation?

  1. The no-drop queue drops packets when it is full
  2. Pause frames are transmitted continuously regardless of queue depth
  3. The port resets each time the queue fills
  4. Pause frames are transmitted when the no-drop queue crosses its threshold and the peer honors them Correct answer.

Explanation

Correct PFC behavior means pause frames are sent only when the no-drop queue exceeds its threshold and the peer responds by pausing transmission. Continuous pause frames or drops in a no-drop queue indicate misconfiguration, and port resets are not expected behavior.

Question 44AI Infrastructure

Which consideration is most important when choosing a routing protocol for the underlay of a large AI fabric?

  1. Compatibility with PFC pause frames
  2. Fast convergence and ECMP support at scale Correct answer.
  3. Ability to carry EVPN routes
  4. Support for VXLAN encapsulation

Explanation

The underlay must converge quickly and support many equal-cost paths so that large, latency-sensitive workloads are not disrupted. VXLAN encapsulation and EVPN routes belong to the overlay, and PFC is independent of the routing protocol.

Question 45AI Infrastructure

A team wants to deploy a two-tier spine-leaf topology instead of a three-tier core-aggregation-access design for a GPU cluster. What is the primary benefit?

  1. Elimination of the need for routing protocols
  2. Support for larger Layer 2 broadcast domains
  3. Deterministic, low-latency paths with consistent hop count between any two leaves Correct answer.
  4. Reduced need for redundant power

Explanation

A spine-leaf design gives every leaf the same number of hops to every other leaf, which produces predictable latency and simplifies ECMP. It does not remove the need for routing, and it typically reduces rather than expands broadcast domains.

Question 46AI Infrastructure

An operator needs to confirm that a no-drop class is correctly configured on a leaf interface for RoCEv2. Which configuration element is essential?

  1. ECN disabled on the class
  2. PFC enabled on the class with a matching CoS-to-queue mapping Correct answer.
  3. A larger MAC address table
  4. A lower MTU on the interface

Explanation

A no-drop class requires PFC enabled for the class and a CoS-to-queue mapping that matches the marking used by the servers. Disabling ECN, enlarging the MAC table, or lowering MTU does not create lossless behavior.

Question 47AI Infrastructure

A fabric team notices that ECN marking is occurring but GPU applications still report retransmissions. Which explanation is most plausible?

  1. The underlay MTU is too large
  2. PFC is disabled on the spine uplinks
  3. The end hosts do not implement congestion control responses to ECN marks Correct answer.
  4. The spine switches are dropping all ECN-marked packets

Explanation

ECN only helps if the endpoints react to congestion notifications by reducing their sending rate; if hosts ignore the marks, congestion and retransmissions continue. Spines dropping marked packets, disabled PFC, or excessive MTU do not match the described symptoms.

Question 48AI Infrastructure

Which design choice most reduces the risk of a PFC storm propagating across an AI fabric?

  1. Scoping no-drop behavior to the specific class and limiting pause propagation at fabric boundaries Correct answer.
  2. Disabling ECN on all leaf ports
  3. Increasing the pause quanta to maximum on all ports
  4. Enabling PFC on all classes on every port

Explanation

Restricting PFC to the class that carries RDMA and preventing pause propagation beyond the intended domain limits the blast radius of a PFC storm. Enabling PFC broadly, maximizing pause quanta, or disabling ECN increases the risk of widespread pause propagation.

Question 49AI Infrastructure

A network architect must choose between PFC and ECN as the primary congestion control mechanism for a RoCEv2 fabric. Which reasoning is most sound?

  1. Use ECN only on spines and PFC only on servers
  2. Use PFC as the primary mechanism and disable ECN
  3. Use neither, relying on TCP retransmission
  4. Use ECN as the primary signal and PFC as a last-resort safety net Correct answer.

Explanation

ECN provides early congestion feedback that lets endpoints slow down before buffers overflow, while PFC prevents drops when buffers are nearly exhausted. Making PFC primary or removing ECN increases pause storms and head-of-line blocking, and relying on TCP is not applicable to RoCEv2.

Question 50AI Infrastructure

An AI fabric is being scaled from 128 to 512 GPU nodes. Which change is most likely to require redesign of the underlay addressing and ECMP scheme?

  1. Adding more MAC addresses to the leaf tables
  2. Moving from a two-tier to a three-tier spine-super-spine topology Correct answer.
  3. Enabling PTP on all switches
  4. Increasing the number of VLANs on each leaf

Explanation

Beyond the port scale of a two-tier fabric, a super-spine tier is added, which changes the addressing plan and ECMP path diversity across the fabric. VLAN count, MAC scale, and time synchronization do not fundamentally alter the underlay addressing or ECMP design.