Design Principles and Future Trends of Data Center Networks

1. Data Center Network Architecture Design Principles

Network is the most critical component of a data center, primarily composed of numerous layer‑2 access devices and a few layer‑3 devices. Early data centers used simple interconnections of a few dozen devices, but modern data centers require higher performance, reliability, and adaptability.

Scalability

To support business growth, evolving demands, and advanced technologies, the network must be modular, employ high‑density ports, and provide routing capabilities at each layer, ensuring strong routing scalability and the ability to deliver value‑added services.

Availability

Redundant design for both equipment and the network itself is essential. Critical devices use carrier‑grade dual redundancy, each layer adopts a dual‑device configuration, and full‑mesh connections between layers provide multiple redundancy options.

Flexibility

The network should be customizable to meet diverse user requirements, offering various common interfaces and allowing logical combination of network modules.

Security

Security, a top concern for data center users, includes both physical space controls and network security measures.

2. Data Center Network Topology Design

Fabric Network

With the rise of cloud computing, server virtualization is widely adopted. To ensure seamless VM migration without service interruption, the underlying layer‑2 network must be sufficiently large, as VM migration is limited to the same layer‑2 domain.

Overlay Network

Overlay adds a virtualized layer on top of the physical network, enabling application‑level services without extensive changes to the underlying infrastructure. It encapsulates traffic in IP packets, allowing routing across the network without special hardware constraints, and supports large‑scale expansion, fault‑tolerance, and load balancing.

Overlay addresses three major challenges:

Enables VM migration across broader network domains by using routing rather than layer‑2 constraints.

Reduces MAC address requirements on access switches, allowing many more VMs per physical port.

Introduces a 12‑bit VLAN‑like identifier (Tenant ID) supporting millions of tenants, eliminating VLAN‑based traffic waste.

Spine‑Leaf Network

The spine‑leaf architecture, derived from the CLOS network, provides a flat, non‑blocking topology with full cross‑connections between spine and leaf switches. This design offers high reliability—failure of a single switch does not affect the overall network—and scalable bandwidth as more leaf switches are added.

BGP EVPN

BGP EVPN, based on the BGP protocol, requires deployment on network switches acting as VTEP nodes to perform VXLAN encapsulation. Switch interfaces or VLANs map to broadcast domains (BD) bound to EVPN instances, enabling VXLAN tunnel creation, MAC learning, and inter‑data‑center connectivity through BGP‑EVPN‑established tunnels.

3. Future Trends of Data Center Networks

Data center networks are evolving from data‑centric to compute‑centric architectures. Key trends include:

Accelerated bandwidth development : Separation of management, control, and data planes with software‑defined networking to build high‑performance, low‑latency networks.

High‑density heterogeneous compute clusters : Shift toward I/O‑centric designs offering elastic scaling, low cost, and high reliability.

Cost reduction and intelligent visualized operations : Adoption of single‑chip box devices, energy‑efficient designs, and automated, AI‑driven operation and maintenance.