Use Case Index

EVPN-VXLAN in a Data Center

Traditional data centers used to have chassis-based switches that were difficult to manage. With changing organizational requirements, data center fabrics are also changing. Nowadays, data centers are expected to support virtualization, span multiple geographies, incorporate hybrid cloud elements, and provide infrastructure for AI workloads. Your data center can meet these requirements by implementing a 3-stage Ethernet VPN-Virtual Extensible LAN (EVPN-VXLAN) fabric architecture with Juniper Apstra. This architecture uses edge-routed bridging (ERB) that distributes traditional network chassis into a switching fabric.

With the QFX5220, QFX5230, QFX5240, and QFX5241 Switches running Junos OS Evolved and using Juniper Apstra as an orchestration platform, you can now provision, manage, and monitor your data centers. The QFX5220 Switch works as a spine transit node in the Ethernet Routing Bridge (ERB) or Border Overlay (BO) design to provide transit connectivity between leaf devices and to aggregate traffic toward the border or edge node. The QFX5230, QFX5240, and QFX5241 Switches support server leaf nodes. These platforms provide the necessary capabilities for leaf-level VXLAN termination and EVPN control plane participation in this architecture.

For more information about the 3-stage EVPN-VXLAN fabric architecture, see 3-Stage Data Center Design with Juniper Apstra (JVD).

To configure EVPN-VXLAN in your data center environment, see EVPN VXLAN Configuration.

Collapsed Spine Switches

A collapsed-spine architecture in a data center combines the functions of spine and leaf switches into a single layer of devices. This architecture thereby eliminates the leaf layer, collapsing its functions onto the spine switches. The collapsed spine switches perform underlay-routing and EVPN-VXLAN overlay functions.

An enterprise that doesn't require a 3-stage data center network might opt for collapsed spine switches in its data center. Deployment scenarios include small, high-availability data center networks, single-rack pods within a larger data center, remote sites and branch offices, and deployments with low budget, space, or power constraints.

For more information about this use case, see Collapsed Data Center Fabric with Juniper Apstra—Juniper Validated Design (JVD).

To configure this use case in your data center environment, see Collapsed Spine Switches Configuration.

AI Cluster Networking

To meet the bandwidth requirement of artificial intelligence–machine learning (AI-ML) workloads and storage systems, you can use QFX5230 and higher-numbered switches. Most switches offer 51.2-Tbps throughput with 800GbE, 400GbE, 200GbE, 100GbE, and 50GbE interfaces and high port density, supporting intra-IP fabric and network interface card (NIC) connectivity for AI-ML use cases. These switches support remote direct memory access (RDMA) and Remote Direct Memory Access over Converged Ethernet version 2 (ROCEv2) with congestion management features such as priority-based flow control (PFC), explicit congestion notification (ECN), and data center quantized congestion notification (DCQCN). All these features are useful in processing AI-ML workloads that transfer data at the network layer.

For information about this use case, see AI Data Center Network with Juniper Apstra, Nvidia GPUs, ConnectX NIC, and Weka Storage—Juniper Validated Design (JVD).

EVPN-VXLAN in AI Cluster Networking

The QFX5230 and higher-numbered switches can have EVPN-VXLAN as an overlay over a pure IP fabric. This setup supports leaf, spine, and super-spine deployments for AI data centers. This cluster supports either a 3-stage IP fabric architecture or a 3-stage EVPN-VXLAN fabric architecture.

For information about how to build an AI data center, see AI Data Center Network with Juniper Apstra, AMD GPUs, Broadcom Thor2 NIC, AMD Pollara NIC, and Vast Storage—Juniper Validated Design (JVD).