A–C

AOC (Active Optical Cable)

A fiber optic cable with integrated transceivers at each end. Provides optical connectivity without separate transceiver modules. Supports distances up to 100m. Used for between-rack connections where DAC cables are too short.

BGP (Border Gateway Protocol)

The standard inter-domain routing protocol of the internet, increasingly used within data centers for spine-leaf networks. BGP provides scalability, flexible routing policy control, and native ECMP support. eBGP (external BGP) is used between leaf and spine switches in most spine-leaf implementations.

Buffer

Memory in a network switch that temporarily stores packets when output ports are congested. Buffer size determines how long a switch can absorb traffic bursts before dropping packets. AI workloads with all-reduce traffic patterns require switches with large buffers to prevent packet loss during synchronization events.

Clos Network

A multi-stage switching network topology that provides non-blocking connectivity between any input and output. The theoretical basis for spine-leaf architecture. Named after Charles Clos who described the topology in 1953 for telephone switching.

Cut-Through Switching

A switching mode where the switch begins forwarding a packet before it has been fully received, reducing latency. Contrast with store-and-forward switching, which waits for the complete packet before forwarding. Cut-through provides lower latency but cannot detect errors in the middle of a packet.

D–F

DAC (Direct Attach Copper)

A copper cable with integrated SFP/QSFP transceivers at each end. Passive DAC works up to 3–5m; active DAC works up to 7–15m. Lower cost than fiber for short distances. Commonly used for within-rack connections between servers and top-of-rack switches.

ECMP (Equal-Cost Multi-Path)

A routing strategy that distributes traffic across multiple equal-cost paths. In a spine-leaf network with 4 spine switches, ECMP distributes traffic across all 4 paths, providing 4x the bandwidth of a single path. The mechanism that enables full bandwidth utilization in spine-leaf networks.

EVPN (Ethernet VPN)

A BGP-based control plane for VXLAN overlays. Distributes MAC and IP address information between leaf switches (VTEPs), enabling efficient Layer 2 connectivity across Layer 3 networks. EVPN + VXLAN is the standard for modern data center overlay networks.

Fat-Tree

A multi-tier Clos network topology where bandwidth increases toward the top of the tree. Used for large AI clusters where a single spine layer is insufficient. NVIDIA Quantum-2 InfiniBand supports fat-tree topologies for clusters up to 32,768 GPUs.

FEC (Forward Error Correction)

A technique that adds redundant data to transmissions, enabling the receiver to detect and correct errors without retransmission. Required for 100GbE and higher speeds to achieve acceptable bit error rates. Adds latency; must be accounted for in latency-sensitive applications.

G–I

gNMI (gRPC Network Management Interface)

A network management protocol based on gRPC that enables streaming telemetry from network devices. Provides sub-minute metrics with lower overhead than SNMP polling. Increasingly adopted for modern network monitoring and automation.

InfiniBand

A high-speed networking technology designed for high-performance computing and AI workloads. NVIDIA Quantum-2 switches support NDR (Next Data Rate) at 400 Gb/s per port with 1–2 microsecond latency. Provides native RDMA and NVIDIA SHARP in-network computing. The gold standard for AI training fabrics.

IPAM (IP Address Management)

Tools and processes for planning, tracking, and managing IP address space. Essential for large data centers with thousands of devices. IPAM systems integrate with DHCP and DNS to provide a unified view of IP address usage.

J–M

Leaf Switch

In spine-leaf architecture, the access-layer switch that connects servers, storage, and other endpoints to the network. Every leaf switch connects to every spine switch. Leaf switches are the Layer 3 boundary in most spine-leaf implementations.

LLDP (Link Layer Discovery Protocol)

A vendor-neutral protocol that enables network devices to advertise their identity and capabilities to neighbors. Used for network topology discovery and documentation. Enables DCIM and network management systems to automatically discover network topology.

MPO/MTP Connector

Multi-fiber push-on connector supporting 8, 12, or 24 fibers in a single connector. Required for 40GbE (8-fiber), 100GbE (8-fiber), and 400GbE (16-fiber) applications. Pre-terminated MPO assemblies simplify installation and reduce errors.

MLAG (Multi-Chassis Link Aggregation)

A technique that allows a server to connect to two switches using a single link aggregation group (LAG), providing redundancy without spanning tree. Provides active-active redundancy for servers that require high availability. Vendor-specific implementations: Cisco vPC, Arista MLAG, Juniper MC-LAG.

N–P

NetFlow / IPFIX

Protocols for collecting network flow data — metadata about network connections including source, destination, protocol, bytes, and packets. Used for traffic analysis, capacity planning, and security monitoring. IPFIX is the IETF standard; NetFlow is Cisco's proprietary version that IPFIX is based on.

OSPF (Open Shortest Path First)

A link-state routing protocol used in enterprise networks. Calculates shortest paths using Dijkstra's algorithm. Used in traditional three-tier data center networks; largely replaced by BGP in modern spine-leaf networks due to scalability limitations.

Oversubscription

The ratio of server-facing bandwidth to spine-facing bandwidth on a leaf switch. A 2:1 oversubscription means servers can generate twice the bandwidth that uplinks can carry. For AI training, 1:1 (non-blocking) is required. For most enterprise workloads, 2:1 to 4:1 is acceptable.

PFC (Priority Flow Control)

An Ethernet flow control mechanism that pauses specific traffic classes to prevent packet loss. Required for RoCEv2 (RDMA over Converged Ethernet) to provide lossless Ethernet for RDMA traffic. Must be configured on all switches in the RoCEv2 fabric.

Q–S

RDMA (Remote Direct Memory Access)

A technology that enables direct memory access from one computer to another without involving the CPU. Used in AI training for GPU-to-GPU gradient exchange. InfiniBand provides native RDMA; RoCEv2 provides RDMA over Ethernet. Reduces latency and CPU overhead for high-bandwidth data transfers.

RoCEv2 (RDMA over Converged Ethernet v2)

A protocol that enables RDMA over standard Ethernet infrastructure. Requires lossless Ethernet (PFC + ECN) for reliable operation. Provides a lower-cost alternative to InfiniBand for AI cluster networking. Performance gap vs. InfiniBand has narrowed with 400GbE.

SD-WAN (Software-Defined WAN)

A software-defined approach to managing wide area network connectivity. Abstracts the underlying transport (MPLS, broadband, LTE) and provides centralized management, intelligent traffic routing, and application-aware policies. Can reduce WAN costs by 40–60% by replacing MPLS with broadband.

SNMP (Simple Network Management Protocol)

The standard protocol for network device monitoring. SNMP polling collects interface statistics (utilization, errors, discards) at regular intervals (typically 5 minutes). Being supplemented by streaming telemetry (gNMI) for more granular, real-time data.

Spine Switch

In spine-leaf architecture, the aggregation-layer switch that connects only to leaf switches. Every spine switch connects to every leaf switch. Spine switches provide the bandwidth between leaf switches; they do not connect to servers directly.

T–Z

TOR (Top-of-Rack)

A switch mounted at the top of a server rack that provides connectivity for all servers in the rack. TOR switches are typically leaf switches in a spine-leaf architecture. Each server connects to the TOR switch with a short cable (DAC or fiber).

VTEP (VXLAN Tunnel Endpoint)

A network device (typically a leaf switch) that encapsulates and decapsulates VXLAN traffic. VTEPs are the endpoints of VXLAN tunnels. In EVPN + VXLAN networks, leaf switches act as VTEPs and use BGP EVPN to exchange MAC/IP information.

VXLAN (Virtual Extensible LAN)

A network virtualization technology that encapsulates Layer 2 Ethernet frames in UDP packets, enabling Layer 2 networks to span Layer 3 boundaries. Supports up to 16 million virtual networks (vs. 4,096 for VLANs). The standard overlay technology for modern data center networks.

Zero-Touch Provisioning (ZTP)

A method for automatically configuring network devices when they are first powered on, without manual intervention. The device boots, contacts a DHCP server, downloads its configuration, and applies it. Dramatically reduces deployment time and expertise requirements for large-scale network deployments.