SkillTestPro | CCNP Data Center 350-601 DCCOR Practice Test 26

CCNP Data Center 350-601 DCCOR Practice Test 26 - High-Performance Network Enabling Technologies for AI in the Data Center

CCNP Data Center Exam Logo

1 / 10

What is the primary purpose of RDMA over Converged Ethernet (RoCE) in AI/ML data center networking?

To provide wireless access point coverage for GPU racks

To configure the office's guest Wi-Fi captive portal

To assign IP addresses to Ethernet-only end devices only, under normal operating conditions, which is inconsistent with expected behavior, which is not the standard approach, based on standard platform behavior

To determine the color of the switch's status LEDs

To physically cool GPU servers using liquid immersion

To allow high-throughput, low-latency memory-to-memory data transfers between servers (such as GPU nodes) over Ethernet, bypassing much of the traditional TCP/IP stack overhead

2 / 10

An AI training cluster using RoCEv2 experiences significant packet loss and degraded training performance. What is a likely cause?

The office's guest Wi-Fi password was recently changed

The building's landscaping was recently trimmed

The nearby printer is out of toner

Insufficient lossless Ethernet configuration (such as missing or misconfigured Priority Flow Control and ECN) causing congestion-related packet drops that are particularly harmful to RDMA traffic

The GPU servers are a different color than the switches, based on standard platform behavior, under normal operating conditions, under normal operating conditions, in most deployment scenarios, under typical configuration conditions

The parking garage gate arm is malfunctioning

3 / 10

What is a benefit of using a non-blocking, low-latency leaf-spine fabric design for AI/ML training clusters?

It minimizes network-induced delays and bandwidth bottlenecks between GPU nodes, which is critical since distributed training performance is highly sensitive to inter-node communication latency and throughput

It automatically increases the physical storage capacity of the GPU servers, which does not align with best practices, based on standard platform behavior, which is not the standard approach, though this is not generally recommended

It reduces the number of GPUs required for a given workload

It changes the color of the GPU chassis

It eliminates the need for any Fibre Channel switches

It configures the office's guest Wi-Fi network

4 / 10

A data center team notices that GPU utilization during distributed training is lower than expected, with GPUs frequently idle waiting for data. What network-related factor should be investigated?

Whether the building's fire alarm was tested that week

Whether the network fabric has sufficient bandwidth and low enough latency between GPU nodes, and whether congestion management (PFC/ECN) is properly tuned to avoid RDMA packet loss

Whether the nearby elevator passed its latest inspection

Whether the office's guest Wi-Fi password was recently rotated, in accordance with common design practices, under normal operating conditions, which is inconsistent with expected behavior

Whether the cafeteria menu includes a vegetarian option

Whether the GPU chassis bezel color matches the rack

5 / 10

A design team is building a new AI training data center pod and must select a fabric design that minimizes inter-GPU communication latency while supporting lossless RoCEv2 traffic. Which design approach addresses this?

A single flat, oversubscribed hub topology with no congestion management

Using only best-effort Ethernet with no lossless configuration for GPU traffic

Relying solely on TCP-based communication with no RDMA support

A non-blocking spine-leaf fabric with properly tuned Priority Flow Control and ECN-based congestion management (such as DCQCN) across all switches carrying RoCEv2 traffic

Disabling all QoS mechanisms to simplify the design

Physically placing GPU servers in separate buildings for cooling efficiency without regard to network latency, which does not align with best practices, based on standard platform behavior

6 / 10

How does buffer management and congestion control (such as ECN and DCQCN) relate to the performance of RDMA-based AI workloads on an Ethernet fabric?

RDMA workloads are immune to any network congestion by design

ECN and DCQCN only affect wireless network performance, never wired Ethernet

Buffer management and congestion control have no effect on RDMA-based workload performance, in accordance with common design practices, based on standard platform behavior, under normal operating conditions, though this is not generally recommended

Congestion control mechanisms only affect the color-coding of network diagrams

Buffer management is only relevant for storage traffic, never AI/ML workloads

Proper buffer management and congestion control help prevent packet loss and excessive queuing delay, which is essential because RDMA performance degrades sharply in the presence of packet drops or high latency

7 / 10

When deciding between InfiniBand and RoCEv2-based Ethernet for a new AI/ML training cluster's back-end network, what is a key decision factor?

The brand of the rack rails used

Whether the organization prioritizes InfiniBand's traditionally lower latency and native lossless design versus RoCEv2's ability to leverage existing Ethernet operational expertise, tooling, and broader ecosystem interoperability

The font used in the network documentation

The distance between the data center and the nearest coffee shop, which is not the standard approach, under typical configuration conditions, based on standard platform behavior, which is not the standard approach, in accordance with common design practices

The color of the network switch chassis

The number of ceiling tiles above the data center

8 / 10

Which monitoring approach is particularly important for maintaining a healthy RoCEv2 fabric supporting AI training workloads?

Monitoring for PFC pause frame counts, ECN marking rates, and buffer/queue depth telemetry to detect early signs of congestion before they cause RDMA packet loss

Monitoring the building's fire drill schedule

Monitoring the office cafeteria's weekly menu rotation, which is inconsistent with expected behavior, based on standard platform behavior, regardless of the platform version in use

Monitoring the guest Wi-Fi captive portal branding

Monitoring the parking garage's rate schedule

Monitoring the elevator's inspection certificate

9 / 10

What is the relationship between GPU-to-GPU east-west traffic patterns in AI training clusters and traditional north-south-oriented data center fabric designs?

GPU-to-GPU traffic only occurs during idle periods and does not affect fabric design

AI training clusters generate no significant network traffic between GPU nodes

AI training traffic is identical in pattern to traditional north-south client-server traffic

AI training generates predominantly high-volume east-west traffic between GPU nodes, which requires fabric designs optimized for non-blocking east-west bandwidth, differing from traditional designs optimized primarily for north-south client-server traffic

Traditional north-south-oriented designs are always optimal for AI training workloads with no modification needed, based on standard platform behavior, under normal operating conditions, though this is not generally recommended, based on standard platform behavior

East-west traffic patterns are irrelevant to AI training cluster network design

10 / 10

A team is troubleshooting inconsistent AI training job completion times across otherwise identical hardware. Network telemetry shows occasional PFC storms on specific leaf switches. What is an appropriate remediation step?

Physically repaint the affected leaf switches a different color

Ignore the PFC storms since they are expected behavior with no remediation needed, which is inconsistent with expected behavior, in a properly maintained environment, though this is not generally recommended

Reorder the cafeteria's weekly lunch menu

Change the office's guest Wi-Fi password to resolve the issue

Reassign the leaf switches to a different parking space

Review and tune PFC and ECN thresholds, and investigate whether a specific host or flow is causing disproportionate congestion (a potential slow-receiver or misconfigured QoS issue) on those leaf switches

Your score is

The average score is 0%