Free NCP-AIN practice test: 30 questions with answers

Per-port utilization and queue occupancy show whether a specific uplink is congested. Training accuracy, GPU core counts and warranty dates do not identify a congested link.

Thresholds with alerts on loss counters give early warning during the run. Checking after failure, disabling counters and relying on user reports detect problems too late.

RDMA bandwidth tests measure achievable throughput over the RDMA fabric between nodes, which is the relevant check. Browser speed tests, DNS lookups and compression tools do not measure the RDMA path.

CRC errors usually point to the physical layer, so checking the cable and optic is the logical first step. Firmware versions, DNS suffixes and hostname lengths do not produce CRC errors.

Applying a validated template and checking it against the baseline keeps the new device consistent. Manual configuration from memory, factory defaults and copying a home router's configuration create inconsistencies.

Bandwidth and latency tests between representative nodes and checks of consistent MTU and flow control along the path reveal problems before a run. Starting the job first, skipping validation and disabling telemetry all make failures harder to prevent or diagnose.

Collective operations are sensitive to latency variation and loss, because every GPU waits on the slowest transfer. Shared unrelated services, panel labels and internet route tables do not determine collective performance.

A dedicated compute fabric keeps GPU traffic free from interference by other traffic types and allows tuning for collectives. Converged networks, management networks and internet paths do not meet the performance and isolation needs.

A balanced leaf-spine fabric gives uniform hop counts and non-blocking capacity, which keeps performance predictable. Unbalanced rings, a single switch with no scaling and WAN hub designs do not provide that property.

A 1 to 1 ratio means uplink capacity matches server-facing capacity, so the fabric is non-blocking. A 10 to 1 ratio is oversubscribed, a single link has no redundancy, and a broadcast domain without uplinks cannot connect the fabric.

Latency, jitter, packet loss and retransmissions directly affect how long collective operations take. Rack door colours, licence names and printers on the subnet do not affect fabric performance.

GPUDirect RDMA moves data directly between GPU and network memory, which reduces copies and latency. The other options do not describe the function of GPUDirect RDMA.

RoCE carries RDMA over Ethernet and requires lossless behaviour and congestion control, which are configured on the switches and NICs. Standard TCP, ICMP and ARP are not RDMA transports.

The InfiniBand subnet manager discovers the topology, assigns local identifiers and computes forwarding paths. DNS, NTP and printer queues do not manage an InfiniBand subnet.

PFC pauses the RDMA traffic class before buffers overflow, which makes the traffic lossless. Disabling flow control allows drops, and DNS TTLs and hostnames have no effect on drops.

ECN marks packets when queues build up, and the sender uses congestion control to slow down before loss occurs. Sending more packets, disabling RDMA and rerouting through management networks do not address congestion.

Consistent priority and flow control settings across all devices, and PFC on the RDMA class, are needed for lossless behaviour. Mismatched MTU values cause problems, disabling congestion notification removes a control, and per-packet traffic classes are not a design requirement.

The LID is the subnet-local address that switches use for forwarding in InfiniBand. Domain names, MAC table entries and VLAN tags belong to other network technologies.

Spectrum-X combines Ethernet switches with adaptive routing and congestion control designed for AI traffic. Default switches lack that tuning, hubs cannot switch traffic at scale, and serial links do not provide cluster-scale fabrics.

Segmentation with VLAN or VXLAN overlays and tenant policies separates traffic and enables isolation. Shared address spaces, a single unclassified queue and open port access remove isolation between tenants.

Streaming telemetry provides real-time counters for utilization, congestion and errors, which supports fast troubleshooting. Monthly reports, weekly manual logins and printed diagrams are too slow or static.

High-bandwidth server links at 400 or 800 Gb/s are common in AI fabrics to move large model data. The lower speeds listed are far below what AI clusters need.

Adaptive routing can spread flows across equal-cost paths and balance load across spines, which reduces hot spots. It does not require a single fixed path, does not remove congestion control, and is used on the data fabric rather than only on management networks.

Adaptive routing with congestion-aware path selection chooses paths based on load, which avoids congested spines automatically. Static routes, spanning tree blocking and yearly manual changes do not respond to congestion.

ACLs and policies restrict traffic between segments and limit what a compromised node can reach. Open reachability, disabled firewalls and shared administrator passwords enable lateral movement.

Automated validation catches policy errors before they reach production and prevents accidental cross-tenant exposure. Shared VLANs, disabled logging and open ports all increase the chance of exposure.

An out-of-band management network isolates management traffic from tenant data and restricts access to administrators. Managing through tenant VLANs, root access for tenants and public internet management all expose the control plane.

Unique strong credentials with centralized authentication remove the widely known defaults, and disabling unused services reduces exposure. Keeping defaults, sharing them and making the interface public all increase risk.

Quotas and rate limits contain the impact of a compromised tenant, and logging with alerting reveals abnormal cross-segment flows. Unlimited access, shared credentials and disabled access lists all widen the impact.

Version control with restricted write access and approval keeps configurations auditable and protected. Shared desktops, email and world-writable files expose configurations to unauthorized change.