Glossary

RP-side control dropping PIM registers from unauthorised sources - the rogue-source defence; pair with register rate-limiting - Multicast routing design - PIM, RP, MSDP.

Policy-model SDN fabric; in AI designs the questions are integration and interconnection with other fabrics - Connectivity models and SDN in AI fabrics.

BGP advertising several paths per prefix so backups pre-exist at the receiver - BGP in enterprise design.

The internal-use group block; carve per-scope/per-app ranges and enforce with multicast boundaries - IPv6 makes scope native in the address - Multicast routing design - PIM, RP, MSDP.

Media-interop PTP profile letting Dante/Livewire/Ravenna/Q-SYS islands share one 1588-2008 clock; mandates clock = EF 46 and media = AF41 - PTP timing for AI fabrics.

Hide reachability detail at topology edges to shrink state and failure domains; leak more-specifics only with a reason - Route aggregation and summarization.

Meta's hyperscale building block - a two-stage Clos (shallow rail leaves, deep-buffered modular spine) joined by an oversubscribed aggregation tier that a topology-aware scheduler compensates - Rail-optimized CLOS topology.

ML over telemetry/logs to detect, correlate, and (gated) auto-remediate at scales humans can't watch - AI-assisted operations - AIOps.

Physically isolated AI environment for maximum-sensitivity work; trades agility and model/data refresh pain for containment - AI security design and hosting models - air-gapped, private, public.

Scheduling so slow clients can't monopolise the channel; capacity design, not politeness - RF design - coverage, capacity, high density, voice, location.

Loss measurement by colouring blocks of live traffic and differencing per-colour counters between two points - exact loss, no probe approximation (RFC 9341) - Network management - traditional vs model-driven.

The serial/synchronized fraction of a workload caps parallel speedup - the reason multi-site training syncs less often the farther apart the GPUs are - Bandwidth and capacity planning.

Intel's on-CPU matrix engine making CPU-only inference viable for smaller models/SLMs - a TCO and placement lever - AI-enabling hardware - GPU, DPU, SmartNIC.

The same distributed default gateway on every leaf so hosts move without re-ARPing - Overlays and BGP EVPN.

Shared RP address on multiple routers with MSDP sync for multicast RP redundancy - Multicast routing design - PIM, RP, MSDP.

Active optical cable; pre-terminated fiber to ~30 m, between DAC and pluggables on the reach/cost/power ladder - Optics and cabling selection.

Election of the single multicast forwarder on a shared segment - lowest AD, then metric, then highest IP toward the source; distinct from the DR and the IGMP querier - Multicast routing design - PIM, RP, MSDP.

Signed provider-authorisation records validating the whole BGP AS_PATH - route-leak protection layered on RPKI's origin-only check - Securing routing protocols.

Set by an L1-L2 border in its L1 LSP once it holds another area's LSPs; L1 routers default toward the closest attached router - the totally-stubby-by-default behaviour that L2->L1 leaking exists to fix - IGP selection - OSPF vs IS-IS vs EIGRP.

An SVI stays up until the last L2 port in its VLAN dies - the hidden ~150-200 ms tax of SVI designs vs routed ports - Physical media and Layer 2 convergence.

The bucket over which BFD loss/latency/jitter are computed; SLA compliance is judged across the last multiplier's worth of buckets - the reaction-time dial - SD-WAN design.

802.1BA deterministic-Ethernet bundle - gPTP time + MSRP per-stream admission + 802.1Qav credit-based shaping; every bridge on the path must participate - PTP timing for AI fabrics.

Cisco mesh path selection - per-hop SNR + hop-count "ease" with a ~20% active-path bonus against parent-flapping - WLAN architectures.

The dedicated lossless network carrying GPU-to-GPU collectives; engineered for JCT, kept separate from front-end traffic - Connectivity models and SDN in AI fabrics.

Bytes in flight a path can hold; sets the TCP window a transfer needs - TCP behavior and the network.

All ranks wait at the same point each step; why one straggler stalls the whole job - Parallelism strategies and collective traffic patterns.

Battery energy storage system - DC-native ride-through and renewables buffer; placement (rack/row/centralized) sets fault-isolation granularity vs units to manage - Sustainability - Green AI, power and cooling.

A border advertises its best external path into iBGP even while an iBGP path wins - pre-positioning the PIC Edge backup - Fast convergence techniques.

Sub-second liveness detection decoupled from protocol hellos; pair it with tuned protocols instead of aggressive timers - Fast convergence techniques.

Tags that carry routing policy between domains; the signalling language of intent - BGP in enterprise design.

Pre-installed backup paths so BGP loss repair is constant-time, not per-prefix - Fast convergence techniques.

Capacity across the fabric's worst cut; the honest number behind 'non-blocking' - Bandwidth and capacity planning.

The set of things a single failure or change can take down; hierarchy, aggregation, and segmentation exist to shrink it - Network hierarchy and topologies.

Best Master Clock Algorithm - per-port, per-announce-interval election (Priority1 first, hop count as tie-break) that builds the PTP hierarchy and pins the grandmaster - PTP timing for AI fabrics.

Array-hosted boot LUNs plus profile-held identity make server hardware fungible - replace or repurpose by re-associating the profile - Storage networking - FC, FCoE, iSCSI, and NVMe-oF

Over-hardened, fully automated defense that shatters instead of degrading; antidotes are graceful degradation and human judgment at high-blast-radius points - Security Design.

Absorbs bursts; deep buffers hide incast at latency cost, shallow buffers need ECN discipline - Buffering, microbursts, and congestion signaling.

FC's hop-by-hop permission to transmit - lossless by construction, the opposite philosophy to Ethernet's mark-then-pause - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Broadcast/unknown-unicast/multicast - the flood classes an overlay must replicate or suppress - Overlays and BGP EVPN.

Capex ownership vs cloud opex as a placement lens; utilisation decides - Cost, ROI, and scalability of AI infrastructure.

Call/connection admission control - QoS protects classes from each other, CAC protects a class from itself; off-path counting or on-path RSVP, with RFC 5865 marking admitted voice - QoS design - models and strategy.

NETCONF's staging workspace - edit and validate there, then commit to running; the protocol-level basis of transactional, rollback-safe change - Network management - traditional vs model-driven.

Cross-connect-rich facility hosting the enterprise cloud edge - private on-ramps to every CSP/carrier without being on-prem - WAN transport architecture selection.

Coolant distribution unit - the liquid-cooling interchange (liquid-to-liquid or locational) that lets RDHx and direct-to-chip scale from an existing hall - Sustainability - Green AI, power and cooling.

Periodic training-state saves; cadence sets DR RPO and drives bursty write I/O the storage tier must absorb - Storage and checkpointing for AI.

NVIDIA reference-configuration notation - CPUs, GPUs, NICs, east-west GbE per GPU (e.g. 2-8-9-400); makes the NIC:GPU ratio an explicit design variable - AI-enabling hardware - GPU, DPU, SmartNIC.

Unsupervised detection of distribution shifts in telemetry time series; diagnostic only when filtered by change type and feature-name semantics - AI-assisted operations - AIOps.

Pipeline-gated change - validate, stage, deploy, roll back - applied to configs - CI-CD pipelines for networks (NetDevOps).

Non-blocking multistage fabric; scale out by adding spines/pods rather than up - Rail-optimized CLOS topology.

Sense (telemetry) -> decide (ML/policy) -> act (gated); the automation end-state - Telemetry-driven and closed-loop automation.

Standby or overflow capacity rented on demand; trades capex for egress, lock-in, and sovereignty exposure - Fabric resiliency and failure handling.

The full observability data set - config, metrics, events, logs, traces - that grounds (agentic) AIOps diagnosis - AI-assisted operations - AIOps.

RADIUS Change of Authorization - server-pushed mid-session re-authorisation (quarantine, guest promotion, posture change); what makes NAC dynamic - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

All-reduce/all-gather style synchronized GPU communication; barrier-synchronised, so the slowest path gates everyone - Parallelism strategies and collective traffic patterns.

BGP exist/non-exist logic - advertise a prefix to an upstream only while a watched route is absent; inbound failover without a permanently visible backup - BGP in enterprise design.

Kernel-sharing failure mode that turns one tenant's pod into host-level access to GPUs and RDMA NICs; why containers are a weaker boundary than VMs - Tenant isolation and segmentation in shared AI clusters.

Rate-limit and filter what reaches the route processor; the device's own DoS survival - Infrastructure hardening - control, management, and data plane.

Detect, compute, and reprogram after a change; each stage has its own levers - Factors affecting convergence.

Cooperative key servers - priority-elected primary, no preemption, identical policy and shared RSA keypair, parallel to GMs, dedicated inter-KS link against split-brain - Tunneling technology selection.

The software-ecosystem switching cost that keeps GPU fleets single-vendor; an interoperability design factor, not just procurement - Workload placement, mobility, and vendor lock-in.

Coarse (cheap, few channels) vs dense (amplifiable, many channels) wavelength multiplexing - Optical transport - CWDM vs DWDM.

Passive copper to ~3 m; cheapest, lowest-power, most reliable link type - default within the rack - Optics and cabling selection.

Label data by sensitivity so placement, encryption, and access rules follow policy - Regulations, governance, and AI security policy.

Large datasets attract compute to themselves because moving them costs time, money, and sovereignty exposure - Data sovereignty and data gravity.

RoCEv2 congestion control built on ECN marks with PFC as backstop; the tuning triangle of lossless Ethernet - Lossless fabric - PFC, ECN, DCQCN.

ECN-fraction-proportional TCP - senders back off by the fraction of marked packets (marked on instantaneous queue depth); smooth low-latency DC transport, the TCP cousin of DCQCN - Buffering, microbursts, and congestion signaling.

Detect, divert, scrub or drop volumetric attacks before they saturate the edge - Threat detection and mitigation - DDoS, spoofing, MITM.

Monitor (auth-open, log-only) -> low-impact (open + pre-auth ACL, per-session overrides) or closed (EAP-only pre-auth); staged per network device group - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Drop ND resolution for destinations absent from the binding table - the defense against sparse-/64 neighbor-cache exhaustion - Layer 2 security.

NAD-local collection of DHCP/CDP/LLDP artefacts shipped in RADIUS accounting - edge profiling without SPAN/NetFlow plumbing - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Provider-to-CPE prefix delegation - the requesting router solicits, the delegating router assigns (a /56 = 256 /64s), and the CPE re-advertises the carved /64s in its RAs - IPv6 design and migration.

Per-application local internet breakout at the branch via data policy - the driver for edge or cloud-delivered security - SD-WAN design.

Hub-anchored multipoint overlay with dynamic spoke-to-spoke tunnels - Tunneling technology selection.

Phase 1 = hub-and-spoke only (deliberate central-policy hairpin); Phase 3 = NHRP-driven dynamic spoke-to-spoke plus hub summarization - the choice is whether traffic must transit a central inspection point - Tunneling technology selection.

How fast a dead link/neighbor is noticed - carrier, BFD, or hellos; the first term in every convergence budget - Physical media and Layer 2 convergence.

Dead peer detection - on-demand IPsec liveness probes sized above the routing hold time so routing converges first and DPD backstops a broken peer - Tunneling technology selection.

SmartNIC plus programmable cores and isolation; runs infrastructure services off-host, enabling bare-metal multi-tenancy and a zero-trust boundary at the NIC - AI-enabling hardware - GPU, DPU, SmartNIC.

Checkpoint-restore plus standby capacity; RPO = checkpoint cadence, RTO = restore + re-provision, replication on the front-end/WAN path - Fabric resiliency and failure handling.

Per-hop-behaviour marking in the IP header; the currency of QoS policy - QoS design - models and strategy.

Two fully independent FC fabrics, never ISL-connected; host MPIO fails over across them so fabric-wide state accidents stay contained - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Two fully independent parallel cores; resilience by switching planes, and services staged one plane at a time - Network hierarchy and topologies.

Run IPv4 and IPv6 together as the default migration posture; every design answer should survive both - IPv6 design and migration.

Authenticate user and machine in one exchange; kills the machine-only backdoor - AAA and identity - RADIUS, TACACS+, 802.1X.

Traffic among servers vs in/out of the block; AI inverts the classic ratio - Connectivity models and SDN in AI fabrics.

Equal-cost multipath by flow hashing; great for many small flows, defeated by few elephant flows without better entropy - Load balancing and hashing - ECMP and entropy.

Mark-before-drop congestion signal; the early, gentle lever that keeps PFC pause rare - Lossless fabric - PFC, ECN, DCQCN.

802.11e/WMM contention tilt - per-access-category AIFS/CW/TXOP make priority frames win the medium more often, not always; wireless QoS is statistical - RF design - coverage, capacity, high density, voice, location.

Cloud data-out pricing; the quiet tax on multi-cloud, mobility, and DR designs - Workload placement, mobility, and vendor lock-in.

DUAL with feasible successors - pre-computed loop-free backups as a protocol property - IGP selection - OSPF vs IS-IS vs EIGRP.

Long-lived high-rate flow (AI collectives are the extreme case) that breaks per-flow hashing assumptions - Load balancing in AI fabrics - flow, flowlet, packet spray.

BGP-signalled overlay control plane (MAC/IP in MP-BGP) replacing flood-and-learn; the standard fabric and DCI language - Overlays and BGP EVPN.

Border-gateway hierarchy interconnecting fabrics with failure-domain isolation between sites - Overlays and BGP EVPN.

A child site of a parent SD-Access fabric site scoping L3VN and IP-pool provisioning to a subset of edge nodes; edge-class nodes only - Overlays and BGP EVPN.

Components failing together because they share a resource; find it before it finds you - Layer 2 fault isolation and resiliency.

Fault, configuration, accounting, performance, security - the management coverage checklist - Reference models and migration considerations.

FC frames on lossless Ethernet (PFC/ETS via DCBX, FIP for login); the convergence trade that only earned single-hop scope - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Forward error correction that buys reach/speed at latency and power cost; watch it at 400G+ - Optics and cabling selection.

InfiniBand's in-band congestion pair - forward mark toward the receiver, backward notification to the sender, adaptive marking rates; the IB analogue of ECN/CNP - Lossless fabric - PFC, ECN, DCQCN.

First-hop gateway redundancy (HSRP/VRRP); align its timers and placement with STP root and the L2 model - Layer 2 design fundamentals.

Operator-defined IGP algorithm (metric + constraints) giving every node a per-algo SID/locator - virtual topologies and slices inside one IGP, with TI-LFA staying inside the slice - Segment Routing.

APs switch locally at the branch while managed centrally; survives WAN loss - WLAN architectures.

Per-flow entropy field usable for ECMP hashing without touching L4 - Load balancing and hashing - ECMP and entropy.

Burst-gap boundary where a flow can be re-hashed without reordering; middle ground between flow and packet spray - Load balancing in AI fabrics - flow, flowlet, packet spray.

Fault-managed power - actively monitored, touch-safe higher-voltage DC distribution with fast selective fault isolation; enables 380VDC buses and retrofit power backhaul - Sustainability - Green AI, power and cooling.

Type-5/7 field that redirects external traffic to the true next hop on a shared OSPF segment instead of the ASBR - third-party next hop for externals - Redistribution and route manipulation.

Transport interface isolated in its own VRF so tunnel underlay and overlay routing can never mix; kills recursive routing by construction - Tunneling technology selection.

The ordinary DC network face of an AI cluster - users, APIs, ingest, storage access, inter-DC - kept off the back-end fabric - Connectivity models and SDN in AI fabrics.

Fibre Channel's link-state routing protocol within a fabric - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

VRF-aware router or firewall outside an SD-Access fabric that leaks routes between isolated VNs and shared services (or the GRT); a firewall fusion is the inter-VN policy point - Overlays and BGP EVPN.

Telecom phase/time PTP profile - full on-path support with a T-BC at every hop; behind 4G/5G radio alignment - PTP timing for AI fabrics.

EU data-protection regime; a hard placement and sovereignty constraint, not a checkbox - Regulations, governance, and AI security policy.

Tunnel-less group encryption preserving the original header; any-to-any over private WAN - Tunneling technology selection.

gRPC operational-actions interface - typed RPCs for ping/traceroute, cert/OS lifecycle, reboot; the structured action side that pairs with gNMI's config/telemetry for machine consumers - Network management - traditional vs model-driven.

Error-free application-layer bytes per second - the number requirements should specify; overhead and retransmissions raise throughput without raising it - TCP behavior and the network.

802.1AS L2-only PTP profile in which every bridge participates; the timing plane of AVB/TSN, mutually exclusive with the default profile - PTP timing for AI fabrics.

NIC-to-GPU-memory transfers bypassing host CPU/memory; why the NIC-per-GPU rail design exists - RDMA, RoCE and RoCEv2.

Storage reads DMA'd straight into GPU memory; moves the storage bottleneck into the network design - Storage and checkpointing for AI.

With SSO, helper neighbors suppress reconvergence and keep forwarding while a restarting router restores protocol state - the forward-through-it philosophy - Fast convergence techniques.

Generic tunnel - simple, multiprotocol, no crypto; the duct tape of overlays - Tunneling technology selection.

Efficiency as a first-class objective - model right-sizing, accelerator utilisation, power/cooling design - Sustainability - Green AI, power and cooling.

Steering clients to surviving sites; the inference-tier DR pattern - Fabric resiliency and failure handling.

Deliberate use of the mobility tunnel - a guest WLAN always tunnels to one static (DMZ) anchor controller, containing all guests at a single policy point - Roaming - L2 vs L3 mobility.

Runtime inspection of prompts and responses against security/privacy/safety policy - enforced inline (gateway proxy) or by API verdict; validated first, then enforced - Protecting AI services - WAF, API security, and inference endpoints.

Untrusted-device onboarding with portal/segmentation; identity decides reach - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Each additional truly independent parallel path roughly halves expected concurrent downtime - void under fate sharing or stateful failover - Network Design.

Record encrypted traffic today, break it post-quantum; justifies PQ pre-shared keys (RFC 8784 / MKA extensions) for long-lifetime secrets - MACsec and secure transport.

High-bandwidth on-package GPU memory; its capacity bounds model size per GPU and drives parallelism choices - AI-enabling hardware - GPU, DPU, SmartNIC.

Dual-stack apps race v6 (with a ~300 ms head start) against v4 and take the winner - broken IPv6 is masked as latency, so monitor the v6 path explicitly - IPv6 design and migration.

Storage controller VM per node pools local NVMe into a shared-nothing distributed store - compute, storage, and GPUs scale in single-node increments; the turnkey inference/RAG archetype - AI-enabling hardware - GPU, DPU, SmartNIC.

One paused priority stalling unrelated traffic behind it; the PFC pathology - Lossless fabric - PFC, ECN, DCQCN.

Synchronous collectives stay intra-site; aggregated updates cross sites less often or asynchronously - makes the DCI bandwidth-heavy but latency-tolerant - Bandwidth and capacity planning.

Layers with distinct roles so change and failure stay local; the oldest scaling trick in the book - Network hierarchy and topologies.

US health-data regime driving encryption, audit, and hosting decisions - Regulations, governance, and AI security policy.

Heat recovery generation - roughly 3% of facility load returned as generation; a rounding error at 100 MW, tens of MW (and a siting requirement) at multi-GW scale - Sustainability - Green AI, power and cooling.

Angle-of-arrival location (~1-3 m @50%) via a circular antenna array; same AP density, serving-AP and orientation constraints - RF design - coverage, capacity, high density, voice, location.

Infrastructure defined as versioned, reviewable code; the source of truth moves out of the boxes - Infrastructure as Code - declarative, idempotent, source of truth.

Re-applying the same definition converges to the same state; what makes automation safe to retry - Infrastructure as Code - declarative, idempotent, source of truth.

Constrain L2 multicast flooding to interested ports; v4 and v6 alike - Layer 2 multicast - IGMP and MLD snooping.

Phased, dependency-ordered steps with per-phase verification, rollback triggers, tested fallbacks, and closing acceptance tests - the exam deliverable beside the design - Reference models and migration considerations.

Stitch a PIM tree straight onto an mLDP LSP by carrying the C-(S,G) in the FEC opaque value - no overlay protocol, but per-flow core state - Multicast routing design - PIM, RP, MSDP.

Many-to-one synchronized bursts overrunning one egress queue; the microburst generator in AI and storage traffic - Buffering, microbursts, and congestion signaling.

Credit-based lossless interconnect with peak collective performance and single-vendor gravity - Fabric transport - Ethernet vs InfiniBand vs UEC.

The no-tree core option - one unicast-LSP copy per egress PE plus an mVPN demux label; zero core multicast state, bandwidth paid at the ingress - Multicast routing design - PIM, RP, MSDP.

Synthetic probes (echo, UDP jitter with responder) measuring delay, jitter, loss, and reachability; via object tracking they actuate static/PBR/FHRP failover on measured health - Network management - traditional vs model-driven.

Level-based link-state IGP; TLV extensibility and clean flooding make it the SP and fabric favourite - IGP selection - OSPF vs IS-IS vs EIGRP.

SCSI over TCP - routable block storage needing nothing special from the fabric; design = dedicated segment, jumbo MTU, MPIO (never LACP), CHAP - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

The metric the whole AI fabric serves; network tail latency and stragglers dominate it - Latency and the straggler problem.

GETVPN's two-key split - the KEK protects the key server's rekey messages, the TEK is the shared traffic key; rekeys are pushed before expiry - MACsec and secure transport.

Namespaces + RBAC + network policy as the software isolation layer above MIG and fabric segmentation - Tenant isolation and segmentation in shared AI clusters.

Per-token attention state that balloons LLM inference memory; why serving is memory-bound - AI-ML workloads and their infrastructure impact.

Point-to-point pseudowires over a plain IP core (no MPLS); chosen by transport availability, not service type - MPLS L3VPN and L2VPN.

LDP, RSVP-TE, or BGP+label as the ways labels learn to mean something - MPLS fundamentals.

One negotiation protocol bundling links; MLAG stretches the bundle across two chassis - Layer 2 multipath and switch clustering.

IGP advertises a link at max cost until LDP has labels on it - prevents L3VPN blackholing onto label-less paths - MPLS fundamentals.

Precomputed loop-free backup next hops for sub-100 ms repair; TI-LFA rides segment routing to cover any topology - Fast convergence techniques.

Direct-to-chip or immersion once racks pass air's ~40 kW ceiling; a facility decision that gates GPU density - Sustainability - Green AI, power and cooling.

Locator/ID split overlay; mobility and scale by mapping instead of flooding routes - Overlays and BGP EVPN.

Strict-priority queue inside CBWFQ; keep the sum of all priority queues within ~33% of the link, never WRED it, and govern it with admission control - QoS design - models and strategy.

The outbound-path thumb on the BGP scale, applied domain-wide - BGP in enterprise design.

Ethernet engineered so RoCE sees effectively zero drops - ECN early, PFC last resort, DCQCN tuned - Lossless fabric - PFC, ECN, DCQCN.

Linear-drive and co-packaged optics; big power savings at 800G+ traded against pluggable flexibility - Optics and cabling selection.

MPLS OAM that verifies the label data plane end-to-end and isolates the broken hop - control-plane up does not mean forwarding works - MPLS fundamentals.

Per-SSID (often rotating) locally administered client MACs by default; kills MAC-as-identity - MAB, allow-lists, analytics - pushing identity to certificates/IdP or per-device PSKs - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Line-rate L2 encryption (802.1AE), hop-by-hop; the WAN variant protects provider transport - MACsec and secure transport.

Between management layers, interfaces change in information (aggregation), services, and protocol - the manager-of-managers job that solves N-x-M tool integration - Reference models and migration considerations.

Session guardrail against route leaks/full-table accidents - BGP in enterprise design.

Snoop-cache-re-advertise proxy carrying link-local service discovery across VLANs and turning it into a policy point (service list, origin, LSS proximity) - Layer 2 multicast - IGMP and MLD snooping.

Per-stage TPM digests extended into an audit trail of everything loaded - detects compromised secure-boot signing keys after the fact; verified by remote attestation against golden values - AI-enabling hardware - GPU, DPU, SmartNIC.

Second factor on top of credentials; assumed by zero trust - AAA and identity - RADIUS, TACACS+, 802.1X.

Fraction of the accelerators' peak FLOPS a training job actually achieves - the utilization metric that stragglers and failures erode - Fabric resiliency and failure handling.

Workload-granular policy vs VRF/VN-scale separation; pick per blast-radius need - Infrastructure segmentation.

Sub-second queue spike invisible to average counters; buffered or marked, never wished away - Buffering, microbursts, and congestion signaling.

Transient forwarding loop while nodes converge at different speeds - Factors affecting convergence.

Hardware partitioning of one GPU into isolated instances with dedicated memory/compute; the hard multi-tenant choice vs time-sliced vGPU - AI-enabling hardware - GPU, DPU, SmartNIC.

Hardware-isolated MIG instances presented through the hypervisor's vGPU operational model - predictable QoS plus VM-estate management; note MIG disables GPU peer transfers - AI-enabling hardware - GPU, DPU, SmartNIC.

LACP floor that fails an entire bundle when members drop below it, turning a hidden capacity loss into an event the IGP can route around - Layer 2 multipath and switch clustering.

Two switches presenting one logical LACP peer; kills the blocked uplink without stacking control planes - Layer 2 multipath and switch clustering.

Multipoint LDP - receiver-driven, downstream-only P2MP label signaling over existing LDP sessions; a tree is (type, root, opaque) and a P router keeps one LFIB entry per tree - Multicast routing design - PIM, RP, MSDP.

Splitting a model across GPUs (tensor/pipeline) when it exceeds one GPU's memory; multiplies east-west traffic - Parallelism strategies and collective traffic patterns.

YANG models with NETCONF/gNMI push-pull replacing SNMP polling - Network management - traditional vs model-driven.

Multicast-only Fast ReRoute - the egress joins a tree twice over disjoint IGP paths, forwards one copy and discards the other until failure; needs ECMP toward the root - Multicast routing design - PIM, RP, MSDP.

Any-to-any tree with a configured root; one tree replaces a P2MP full mesh for a default MDT, the root protected by anycast or hot-standby redundancy - Multicast routing design - PIM, RP, MSDP.

Host-side, storage-aware multipathing across independent paths/fabrics; the failover layer the dual-fabric doctrine relies on - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

RFC 4364 control plane whose next hops resolve to BGP-learned IP-encapsulation endpoints - no LDP, no NHRP, no per-peer tunnels; segmentation over an SP IP-VPN that exchanges no labels - Tunneling technology selection.

Inter-domain RP gossip enabling anycast RP and interdomain multicast source discovery - Multicast routing design - PIM, RP, MSDP.

Multisite Remote Border - anchors an SD-Access VN to a different fabric site's border and control plane (the fabric generalisation of the guest anchor); edges VXLAN straight to the remote border - Overlays and BGP EVPN.

Per-stream admission control (802.1Qat) - talker/listener declarations reserve bandwidth plus a worst-case latency bound at every hop, or the stream is refused - PTP timing for AI fabrics.

Admit, profile, and segment endpoints by identity/posture at the port - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Address translation that creates state and breaks end-to-end symmetry assumptions - Generic routing and addressing - PBR, NAT, RIB-FIB.

802.1X supplicant-switch pattern (with CISP) - a small edge switch authenticates to the access layer and its port flips access -> trunk; authenticated capacity beyond the closet - AAA and identity - RADIUS, TACACS+, 802.1X.

Border rewrites eBGP NEXT_HOPs to its loopback so iBGP peers resolve one interior address and edge subnets stay out of the IGP - BGP in enterprise design.

Resolution protocol mapping overlay (tunnel) addresses to transport (NBMA) addresses; registrations, redirects, and shortcuts are DMVPN's real control plane - Tunneling technology selection.

Fabric capacity equal to host capacity at every tier; the AI back-end default - Bandwidth and capacity planning.

L2-adjacent PSN peers heartbeating; a survivor fires CoA to restart a mid-auth session on member death - matters chiefly behind a load balancer - AAA and identity - RADIUS, TACACS+, 802.1X.

N-Port ID Virtualisation - many fabric logins on one physical N-port; how FI-attached vHBAs appear to the SAN - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Non-Stop Routing - protocol state kept refreshed on the standby RP internally, no peer cooperation needed, at higher resource cost than GR - Fast convergence techniques.

Neighbour Unreachability Detection - per-neighbour reachability probing in ND; powers RA-based gateway failover (~40 s at defaults) - IPv6 design and migration.

Intra-server GPU interconnect (scale-up domain); what the network doesn't have to carry - Scale-up vs scale-out.

NVMe semantics across the network (RoCE/TCP/FC transports); parallel storage access at fabric speed - Storage and checkpointing for AI.

On-Demand Next-hop - SR-TE policies instantiated from a template when the first matching colored service route arrives; steering intent as route policy, not per-destination configuration - Segment Routing.

Subcarrier scheduling (Wi-Fi 6+) turning airtime into schedulable resource units - Wireless fundamentals and 802.11 standards.

SD-WAN's overlay control protocol carrying routes, TLOCs, and policy - SD-WAN design.

Observe-Orient-Decide-Act; security design as a race of loops - pre-position visibility, playbooks, and containment to shorten the defender's loop - Security Design.

Model whose weights are published and self-hostable - the only class eligible for air-gapped or sovereignty-bound placements; closed weights mean hosted-API-only consumption - Service placement - on-prem, cloud, hybrid, distributed.

Optical signal-to-noise ratio - the amplified-system design currency; EDFAs add noise as they amplify, and modulation density sets the OSNR a rate demands (400G 16QAM ~22.5 dB vs 100G QPSK ~11.5 dB) - Optical transport - CWDM vs DWDM.

Optimal Route Reflection - the RR computes best-path from each client's IGP vantage (link-state view required), restoring hot-potato from centralised or virtual RRs - BGP in enterprise design.

Area-based link-state IGP; area design is aggregation design - IGP selection - OSPF vs IS-IS vs EIGRP.

Provider-independent overlay you build yourself (DMVPN/IPsec/SD-WAN) over any transport - vs the provider-coordinated MPLS VPN service - WAN transport architecture selection.

Virtual topology tunneled over a simple IP core; state moves to the edges - Overlays and BGP EVPN.

IS-IS LSP flag - the domain stops using the router as transit while its connected destinations stay reachable; on-startup and maintenance drains (OSPFv3 analog: R flag) - IGP selection - OSPF vs IS-IS vs EIGRP.

Ratio of offered load to fabric capacity per tier; AI back-ends run 1:1, campus runs economics - Bandwidth and capacity planning.

Data-lake object store with dual HDFS + S3 APIs; blocks grouped into storage containers, namespace split from storage management; scales past HDFS node ceilings - Storage and checkpointing for AI.

Per-packet load balancing for elephant flows; needs reorder tolerance at the NIC or transport - Load balancing in AI fabrics - flow, flowlet, packet spray.

The mVPN Profile 14 core tree - built on demand, connecting only the PEs with traffic, no always-on default MDT; unidirectional and anycast-source-safe - Multicast routing design - PIM, RP, MSDP.

Forwarding by policy instead of destination; powerful, stateful-feeling, easy to orphan - Generic routing and addressing - PBR, NAT, RIB-FIB.

DMVPN hub applies a per-spoke parent shaper (selected by the spoke's NHRP group at registration) so spokes cannot starve each other; hub-only - spoke-to-spoke tunnels bypass it - Tunneling technology selection.

Per-priority pause making one class lossless; overuse spreads congestion and risks pause storms - the safety net, not the mechanism - Lossless fabric - PFC, ECN, DCQCN.

Detects stuck-pause conditions and breaks the deadlock; the operational guardrail on lossless classes - Lossless fabric - PFC, ECN, DCQCN.

Multicast delivery flavours - shared-tree, bidirectional shared-tree, and source-specific without an RP - Multicast routing design - PIM, RP, MSDP.

Functional module - campus, branch, WAN, DC, edge - each with its own paradigm and policy; decomposition by role, not topology - Network hierarchy and topologies.

Track multicast join/prune per neighbour address on a multipoint (mGRE/DMVPN) interface so one spoke's leave does not prune all; sparse-mode only - Multicast routing design - PIM, RP, MSDP.

A proxy address carried in the PIM join so routers can RPF toward a border (ASBR/ABR) when the real root is invisible to the local IGP - inter-AS, CsC, seamless MPLS - Multicast routing design - PIM, RP, MSDP.

Protected management frames; stops deauth/disassoc spoofing, mandatory with WPA3 - Wireless security.

Path MTU discovery; ICMP filtering breaks it, tunnels shrink it, design for both - MTU, fragmentation, and PMTUD.

Failure -> churn -> CPU load -> missed hellos -> more failure; broken by summarization, flooding boundaries, and dampening - Factors affecting convergence.

Agent or temporal-agent compliance check (firewall/patch/registry/app state) gating authorisation; non-compliant sessions redirect to remediation and CoA promotes on success - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Enforced per-device/rack power ceilings trading peak performance for facility headroom - Sustainability - Green AI, power and cooling.

Holding FHRP preemption until the recovered box is forwarding-ready (L1/L2/L3), avoiding a second black-hole window on recovery - Layer 2 fault isolation and resiliency.

Collective-library mechanism that moves a message across NVLink inside the sending host to the GPU whose NIC sits on the destination's rail before injection - send-side only in NCCL/RCCL, so a dead rail drops rather than re-rails - Rail-optimized CLOS topology.

Preferred Scanning Channels - the every-fourth 6 GHz channels where clients may actively probe; off-PSC hidden WLANs are unfindable - RF design - coverage, capacity, high density, voice, location.

Sub-microsecond time distribution (boundary/transparent clocks); aligns telemetry and time-sensitive workloads across the fabric - PTP timing for AI fabrics.

Isolation inside one subnet - promiscuous, isolated, community ports - Infrastructure segmentation.

Context sharing and SGT transport between policy platforms and enforcement points - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

ISE syncs CMDB records into local dictionaries (correlation/unique/version identifiers) so authorization can match and return asset attributes - segmentation intent sourced from the system of record - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Quality of Experience - the user-side superset of QoS: access, security, app-aware routing, path, segmentation, and visibility composed as one identity-anchored overlay policy - SD-WAN design.

Where markings start being believed; everything outside gets re-marked - QoS design - models and strategy.

Lower-precision weights (8-bit and below) shrinking memory/compute; the lever that moves models down-tier to SLM/CPU/edge - AI-ML workloads and their infrastructure impact.

RDMA's per-conversation send/receive channel and RoCEv2's unit of flow identity; QP count and QP-aware hashing are the ECMP entropy levers - Load balancing in AI fabrics - flow, flowlet, packet spray.

Priority for the latency class, weighted shares for the rest; protect little, protect well - QoS design - models and strategy.

Encrypted UDP transport with streams; invisible to middleboxes, shifts control to endpoints - Transport protocols overview - TCP, UDP, QUIC.

Per-interface RA bits selecting host addressing - A = SLAAC, O = stateless DHCPv6 options, M = stateful DHCPv6; M=1 with A=1 leaves hosts with two GUAs - IPv6 design and migration.

Blocks rogue IPv6 router advertisements at the access port - Layer 2 security.

Retrieval-augmented generation grounding a model on a private knowledge base; often the on-prem answer to sovereignty - AI-ML workloads and their infrastructure impact.

The rails terminating on one leaf in a 2/4/8-way rail design - traffic between rails of the same group is single-hop, even cross-rail, with no spine transit or PXN - Rail-optimized CLOS topology.

NIC-per-GPU rails each on their own leaf plane so same-rank GPUs are one hop apart; the AI-specific CLOS variant - Rail-optimized CLOS topology.

Per-client transactions-per-second budgets with throttling - an orchestration-design constraint that punishes naive polling - Network automation fundamentals - why and the operating model.

Rear-door heat exchanger - locational liquid cooling (~20-40 kW/rack) that fixes hot spots without floor redesign; the brownfield on-ramp to direct-to-chip - Sustainability - Green AI, power and cooling.

GETVPN migration state - members forward cleartext but accept ciphertext until the key server flips the group to full SAs; a single-point network-wide encryption cutover - MACsec and secure transport.

Congestion control moved into the collective library - a sender posts an RDMA write only after the receiver's clear-to-send (CTS), so channel buffers cap in-flight traffic - Lossless fabric - PFC, ECN, DCQCN.

HCI protection by keeping RF copies of each write across nodes - RF2 halves usable capacity and replaces array RAID or erasure stripes - Storage and checkpointing for AI.

OMP attribute pinning a colour to same-colour tunnels only - the off switch for full-mesh-by-default - SD-WAN design.

Control-plane best paths compiled into forwarding hardware; recursion between them is a convergence factor - Generic routing and addressing - PBR, NAT, RIB-FIB.

Reduced Neighbour Report - 2.4/5 GHz beacons advertise the co-located 6 GHz BSS; the primary 6 GHz discovery path - RF design - coverage, capacity, high density, voice, location.

RDMA over routable UDP/IP; brings RDMA to Ethernet but imports the losslessness requirement - RDMA, RoCE and RoCEv2.

A next hop resolved via another lookup; a hidden dependency in convergence - Factors affecting convergence.

L3 to the access layer - no STP blocking, but VLANs stop spanning - Layer 2 design fundamentals.

Cryptographic origin validation for BGP; table-stakes edge hygiene - Securing routing protocols.

Data you accept losing vs time you accept being down; for training, checkpoint cadence and restore time respectively - Fabric resiliency and failure handling.

Per-VLAN rapid spanning tree; blocking becomes a per-VLAN traffic-engineering knob - Spanning Tree - types, tuning, and loop mitigation.

Signalled MPLS tunnels with per-LSP state; the price of explicit paths before SR - MPLS fundamentals.

RT-constrained route distribution - PEs signal RT membership so the RR filters VPN updates per PE; update volume scales with imports, not the table - MPLS L3VPN and L2VPN.

Raised receive threshold making the AP treat weak arrivals as noise - shrinks the uplink cell to shed sticky clients; the high-density receive-side lever - RF design - coverage, capacity, high density, voice, location.

The vendor-neutral programming API between a NOS and the switch ASIC - SONiC's intent-to-silicon translation layer, and the reason BYO-NOS is portable across merchant silicon - Connectivity models and SDN in AI fabrics.

The enterprise growth granule - NVIDIA ERA's 4 HGX nodes (32 GPUs) added as a block (N-S/E-W spine split arriving around 16 nodes); Cisco AI PODs package the leaf pair inside the granule so the leaf port-class sets the cluster quantum - Validated reference design - Cisco AI-ML lossless fabric (CVD).

The third scaling axis - GPU communication between clusters or sites over long-haul, jitter-controlled RDMA DCI - Scale-up vs scale-out.

Bigger node (NVLink domain) vs more nodes (network domain); where the domain boundary sits is the core AI compute decision - Scale-up vs scale-out.

Source-routed paths as label/SID stacks in the IGP; TE without per-flow state in the core - Segment Routing.

Cloudera's shared data experience - a governance/metadata layer independent of compute and storage; policy defined once travels with data across clusters and clouds - AI-enabling hardware - GPU, DPU, SmartNIC.

Identity-derived tags enforcing policy independent of IP topology; segmentation that survives readdressing - Network access control and segmentation - NAC, TrustSec, guest and BYOD.

Old and new control planes coexisting untranslated during migration - Reference models and migration considerations.

Meet-point link joining otherwise parallel core planes - a high-cost alternate path that also couples the planes' failure domains - Network hierarchy and topologies.

Hardware quietly computing wrong results with no alert; defended by gradient-norm watching, prologue/epilogue diagnostics, and burn-in discipline - Fabric resiliency and failure handling.

Small language model - cheaper, lower-latency, edge/CPU-deployable with narrower scope; the counterweight to central LLM clusters - AI-ML workloads and their infrastructure impact.

A device withholding B2B credits, backing congestion into the fabric; watched via tx-wait and credit counters, answered by edge port-guard eviction - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Next-gen device clustering on standard Ethernet - IS-IS SPF control plane + VXLAN-GPE encapsulation, front-panel or dedicated links; the cluster becomes a small routed fabric - Layer 2 multipath and switch clustering.

One update-group member that cannot drain its queue stalls replication for the whole group; detect and quarantine into its own update group - BGP in enterprise design.

NIC with hardware offloads (RDMA, overlay, crypto) freeing host CPU; the DPU adds cores and isolation on top - AI-enabling hardware - GPU, DPU, SmartNIC.

Activation-smoothing quantisation keeping int8 accurate beyond weight-only schemes - the accuracy keeper for full-int8 CPU inference - AI-enabling hardware - GPU, DPU, SmartNIC.

Storage Policy-Based Management - vSphere policies (fed by VASA capability data) that Kubernetes StorageClasses name; the tiering contract on virtualised container platforms - Storage and checkpointing for AI.

DV rule - never advertise a route out the interface it was learned on (poisoned reverse: advertise it as unreachable instead); blocks spoke-to-spoke re-advertisement at multipoint hubs - IGP selection - OSPF vs IS-IS vs EIGRP.

Sparse-mode receivers join the shared tree (source unknown until the first packet) then switch to the source tree; threshold-infinity keeps (*,G)-only state at the cost of path stretch - Multicast routing design - PIM, RP, MSDP.

Shared-risk link groups - logical diversity that dies in the same conduit; buy diversity, verify it - Transport resiliency, diversity, and migration.

Segment Routing Mapping Server - advertises prefix-SIDs on behalf of LDP-only routers so SR islands can reach them during migration; deploy redundantly - Segment Routing.

Storage Virtual Machine - the per-tenant storage isolation grain on shared arrays: own volumes, LIFs, admin, and audit domain - Tenant isolation and segmentation in shared AI clusters.

The one slow GPU/link/flow that stalls a barrier-synchronised job; find-the-straggler is the AI-fabric ops loop - Latency and the straggler problem.

Model-driven push (gNMI/YANG) at sub-second cadence; the sensory input for closed-loop operations - Network management - traditional vs model-driven.

EIGRP query left unanswered past the active timer flushes the neighbor's routes - bounded by stub roles and summarization query boundaries - IGP selection - OSPF vs IS-IS vs EIGRP.

VRF/VN-aware SXP - one connection carries IP-SGT bindings for every virtual network (v4 needs one per VN); a reflector bridges v4-only speakers - Infrastructure segmentation.

With firewalls/NAT at a multihomed edge, path symmetry is a design requirement, not an optimisation - Perimeter security - firewalls and IPS-IDS.

Stateless SYN-flood defence - the server encodes half-open-connection state into a cryptographic ISN and reconstructs it from the returning ACK - Threat detection and mitigation - DDoS, spoofing, MITM.

Command-level device-admin AAA with per-command authorization - AAA and identity - RADIUS, TACACS+, 802.1X.

The p99+ delay that gates synchronized workloads; averages lie, tails rule JCT - Latency and the straggler problem.

Time-based anti-replay - GETVPN's replacement for per-peer counters on a shared group SA, using key-server pseudo-time with no NTP dependency - Tunneling technology selection.

Network-side SYN-flood defence - a device brokers the three-way handshake past an embryonic-connection threshold so only completed handshakes reach the server - Threat detection and mitigation - DDoS, spoofing, MITM.

SLAAC privacy extension - rotating random-IID addresses preferred as source for new connections (RFC 6724 rule 7); the enterprise traceability tension - IPv6 design and migration.

Rarity-weighted term scoring used to rank log lines or features so scarce, informative items surface first - AI-assisted operations - AIOps.

Pointing traffic at a router that didn't advertise the route; a metric-free steering trick - Metric-based traffic engineering.

An edge's transport attachment (system IP + colour + encapsulation) advertised as an OMP route; TLOC validity rides on an active BFD session - SD-WAN design.

Decrypt-inspect-re-encrypt at the perimeter; visibility bought with privacy, latency, and PKI cost - Perimeter security - firewalls and IPS-IDS.

Element -> network -> service -> business management layers; valid element configs can still break the network, so end-to-end integrity needs its own layer - Reference models and migration considerations.

Enterprise-architecture framing that maps business drivers to technical target states - Reference models and migration considerations.

Summarize or tunnel so remote failures don't ripple in; stretch is the price of the hiding - Network hierarchy and topologies.

Inter-token streaming rate of LLM decode - GPU memory-bandwidth-bound; with TTFT it decomposes an inference latency SLA - AI-ML workloads and their infrastructure impact.

Bursty synchronized east-west vs steady latency-bound serving; nearly every AI infra decision forks on this - AI-ML workloads and their infrastructure impact.

Dedicated SD-Access device outside the forwarding path holding aggregate prefixes of every fabric site, so VN and SGT ride VXLAN natively between sites - Overlays and BGP EVPN.

Deliberately elected dual-carrier site that re-advertises between provider clouds (capacity-sized, geo-diverse) while every other site stays non-transit - BGP in enterprise design.

WPA3 clients refuse later WPA2 fallback on an SSID they have seen as WPA3 - closes the mixed-mode downgrade (Dragonblood) door - Wireless security.

Bump-in-the-wire insertion vs being a hop; insertion mode drives the migration plan - Perimeter security - firewalls and IPS-IDS.

Controller-computed (SR-PCE) P2MP multicast trees built from SR MPLS labels - SR's answer to retiring mLDP, bought with a controller dependency - Multicast routing design - PIM, RP, MSDP.

Prompt-to-first-token wait - the compute-bound prefill phase and the responsiveness half of an inference SLA - AI-ML workloads and their infrastructure impact.

Detects one-way fiber before STP believes the lie - Physical media and Layer 2 convergence.

Consortium effort making standard Ethernet an AI-class transport - modern congestion control, multipath, reorder tolerance - Fabric transport - Ethernet vs InfiniBand vs UEC.

Unique local addresses (fd00::/8) with pseudo-random global IDs - private space that merges without renumbering; pairs with NPTv6 for address independence, not security - IPv6 design and migration.

Asymmetric-return pathology where an aged CAM entry turns every downstream packet into a VLAN-wide flood - Layer 2 design fundamentals.

Drop packets whose source fails the reverse-path check; anti-spoofing at the edge, careful with asymmetry - Infrastructure hardening - control, management, and data plane.

Micro-SIDs - 16-bit segments packed under a common block into one IPv6 destination address; a source-routed path transit routers cross with plain longest-prefix match - Segment Routing.

Similarity-search store for embeddings behind RAG retrieval - hot cluster state on block volumes, bulk indexes and logs on S3 object storage - Storage and checkpointing for AI.

Switching cost created by ecosystems (CUDA, IB, proprietary fabrics); designed against with open standards and abstraction layers - Workload placement, mobility, and vendor lock-in.

Time-sliced GPU sharing across VMs; consolidation over isolation, the opposite pole from MIG - AI-enabling hardware - GPU, DPU, SmartNIC.

Per-egress queues held at ingress with scheduled fabric access - kills head-of-line blocking and confines drops to the ingress - Buffering, microbursts, and congestion signaling.

Per-tenant routing table; macro-segmentation's basic unit - Infrastructure segmentation.

Independent virtual fabric (own zoning and FSPF) on shared FC switches - the SAN's VRF - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

VLAN database propagation; modern designs disable it - one bad revision is a campus-wide blast radius - Layer 2 design fundamentals.

UDP-encapsulated L2/L3 overlay with 24-bit segment IDs; the data plane under EVPN fabrics - Overlays and BGP EVPN.

Application-layer filtering in front of AI/inference endpoints - prompt-abuse and OWASP protections live here - Protecting AI services - WAF, API security, and inference endpoints.

Reverse API - the platform POSTs to a registered endpoint on an event, with a validation-token handshake; the push alternative to polling - Network automation fundamentals - why and the operating model.

32-bit metric TLVs replacing the 6-bit narrow style (cost cap 63); incompatible with narrow - a mismatch silently partitions - and prerequisite for multi-topology and L2->L1 leaking - IGP selection - OSPF vs IS-IS vs EIGRP.

Quantize the weights to int8 while activations stay higher-precision - the accuracy-preserving compromise that still engages CPU matrix accelerators for serving - AI-enabling hardware - GPU, DPU, SmartNIC.

TCP option letting windows exceed 64 KB; without it long fat pipes starve - TCP behavior and the network.

Client-side bridge relaying for wired devices behind it; Cisco mode = multi-MAC/VLAN but vendor-locked, universal mode = any-vendor with a single wired MAC - WLAN architectures.

SAE authentication and opportunistic encryption for open networks; the current wireless-security floor - Wireless security.

Early random drop by class to keep TCP queues shallow; the classic cousin of ECN marking - QoS design - models and strategy.

Extended packet numbering (802.1AEbw) - 64-bit MACsec packet numbers so 100G+ links do not rekey-storm; effectively mandatory at speed - MACsec and secure transport.

Never trust by location; authenticate and authorize every session with identity + posture - Zero Trust and ZTNA.

FC access control defining who may talk to whom; single-initiator zoning is the default discipline - Storage networking - FC, FCoE, iSCSI, and NVMe-oF.

Coherent 400/800G pluggables in ordinary switch ports reaching ~120-500 km - metro DCI without a telecom line system - Optical transport - CWDM vs DWDM.

Devices self-provision from the network at first boot; day-0 at fleet scale - Zero-touch provisioning.