Demystifying Clos Fabric: 10 Critical insights into Spine & Leaf Topology
Clos fabric has become the de facto topology for modern, scalable, and application-driven data center designs. Yet, for many engineers transitioning from traditional 3-tier architectures, questions persist around what defines a Clos, what makes it different, and how to use it effectively. This article presents 10 key points that clarify Clos architecture and how it changes the way we think about network design.
1. What is a Clos Fabric?
A Clos fabric is a non-blocking, multi-stage switching topology originally developed for telephone switches, now foundational in modern data center networks. Its primary goal is to scale linearly while offering predictable latency, high bandwidth, and fault tolerance.
Key components:
• Leaf switches connect to servers and endpoints.
• Spine switches act as a fast transit layer between leaves.
• Super spines may exist in very large deployments.
In true Clos fashion:
• Leafs connect to all spines.
• There are no leaf-to-leaf or spine-to-spine connections.
• Any communication between leaves is always two hops (Leaf → Spine → Leaf), though this is an outcome—not a defining rule—of the topology.
Clos fundamentally differs from traditional 3-tier networks where redundancy and loop prevention are bolted on via protocols like STP and FHRP. In Clos, redundancy is integrated inthrough topology and ECMP routing, offering an active-active, deterministic, and protocol-light environment.
2. Traditional Cores vs Spines
In legacy networks:
• The core layer is a central aggregation point, often a single point of failure or bottleneck.
• STP blocks links to prevent loops, wasting expensive bandwidth.
Clos fabric eliminates the “core” concept:
• Multiple spine switches replace the centralized core.
• All leafs connect to all spines, enabling non-blocking, scalable throughput.
• Every link is active, and STP is not required.
This ensures full path utilization, lower latency, and superior east-west performance
3. Spine-to-Spine Connectivity?
A frequent question:
“We used to interconnect cores for redundancy—why don’t we interconnect spines?”
In Clos:
• Spines are never interconnected.
• All traffic follows the Leaf → Spine → Leaf model.
• Spine-to-spine links would violate the deterministic traffic model, introduce loops, and unnecessary complexity.
What about Multi-Pod or Multi-Site?
Yes, spine-to-spine links exist between separate Clos fabrics, using IPN or ISN, to:
• Extend BGP EVPN control planes.
• Carry VXLAN traffic between pods/sites.
But within a single Clos fabric, spine interconnection is intentionally omitted to preserve simplicity and predictability.
4. Traffic Path Reliability
Clos fabrics offer multiple equal-cost paths between any two endpoints via ECMP.
Benefits:
• Loss of a spine or uplink doesn’t impact reachability.
• Traffic automatically reroutes through remaining paths.
• Spines are stateless, allowing fast failover and convergence.
This built-in redundancy and stateless design make Clos ideal for modern workloads demanding high uptime and low latency/Jitter.
5. Multi-Tenancy and Segmentation
Clos fabrics, especially when paired with controllers like Cisco ACI, support true multi-tenancy:
• Use of VRFs, Bridge Domains, and EPGs for segmentation.
• Logical separation without needing to redesign the physical fabric.
• Fine-grained security policies at the application or tenant level.
This allows enterprises to run multiple isolated environments on a shared fabric, with per-tenant policies and services.
6. Clos is a Topology — Not a Solution
A common misconception: Clos alone gives you scale and flexibility.
The truth:
• Clos is the plumbing — the raw topology.
• Without policy and control planes (e.g. Cisco ACI, NSX), you can’t fully leverage its potential without an overlay(e.g VXLAN).
• Think of it as the hardware canvas on which SDN business logic, segmentation, and automation.
7. Where Does Cisco ACI Come In?
ACI is SDN solution by Cisco and a separate entity to Clos, which is in fact a network topology - however
Cisco ACI operationalizes Clos topology by:
• Automating VXLAN/EVPN overlays.
• Centralizing policies, contracts, and application models via APIC.
• Integrating service graphs, firewalls, and endpoint tracking.
ACI transforms Clos from a raw underlay into an intent-driven, programmable fabric, giving you visibility, consistency, and operational simplicity.
⸻
8. Cloud Integration — ACI Anywhere
Modern infrastructure spans multiple clouds. ACI can follow.
ACI Anywhere:
• Extends ACI policy and segmentation to AWS, Azure, and more.
• Offers one dashboard, one model, across on-prem and cloud.
• Reduces operational toil and simplifies hybrid cloud networking.
The Clos design supports this flexibility through its scalable, L3-based ECMP core.
9. Are Public Clouds Using Clos?
Absolutely.
AWS, Azure, and Google Cloud all use Clos or Clos-like topologies:
• Built for horizontal scale, low-latency, and resilience.
• Run massive numbers of ECMP paths.
• Operate overlays for segmentation. We know utilising overlay technologies for multi-tenancy and segmentation is a primary driver for cloud computing.
However, you don’t control this fabric — it’s abstracted behind VPCs/VNets, security groups, and cloud APIs. You rely on the provider’s SDN for performance and security.
10. Clos in the Campus LAN? Yes, You Can.
You can deploy Clos in enterprise LANs, especially in:
• Large campuses
• Dense office buildings
• Data center-to-campus convergence
Why use Clos instead of 3-tier in LAN?
• No STP = no blocked ports.
• ECMP scaling = efficient bandwidth.
• Modular leaf-spine growth = future-ready.
When paired with Cisco DNAC or SD-Access, you gain fabric-wide automation, segmentation, and telemetry — the same principles that made Clos dominant in data centers.
Conclusion: A New Foundation
In traditional networks:
• We interconnect core and distribution switches to patch over limitations in design.
In Clos:
• We design the topology to make those limitations irrelevant.
Clos isn’t just a way to connect switches — it’s a design philosophy: build in scale, redundancy, and simplicity from the ground up.
Redundancy and Scalability are not a separate entity unlike traditional 3 tier network design which needs to be incorporated with subset of protocols such as spanning tree protocol to prevent loops and utilising FHRP to achieve layer 3 redundancy across distribution.
However In a Clos fabric redundancy and scalability are the attributes of the topology it self.
Whether in the data center, in the cloud, or across your enterprise LAN, Clos fabric gives you the architectural freedom to build fast, reliable, and programmable networks for the fut