Architecting the backbone of the cloud. From non-blocking Leaf-Spine topologies and EVPN control-planes to terabit-scale optical fabrics.
Leaf-Spine, Clos & Non-Blocking Design
Deep-dive into dedicated listing pages for every major networking discipline, optimized for professional reference and architectural planning.
Modern datacenter design has pivoted from the legacy 3-tier model to the Leaf-Spine (Clos) topology. By utilizing multiple equal-cost paths (ECMP) between Leaf and Spine switches, we create a non-blocking fabric where any server can communicate at full wire speed. This predictable, low-latency architecture is what enables massive horizontal scaling.
Scaling a shared datacenter requires robust multi-tenancy. VXLAN provides encapsulation to stretch L2 domains across an L3 underlay, while BGP-EVPN serves as the control-plane. This combination eliminates flooding and allows for active-active multi-homing at the edge.
The move to 400G and 800G Ethernet requires highly precise optical fabrics. Engineers must manage sensitive Bit Error Rates (BER) and specialized SerDes configs. At these frequencies, the physical layer becomes a game of signal integrity and thermal management.
In a modern fabric, the default gateway shouldn't be a single physical box. Anycast Gateways allow every Leaf switch to share the same IP/MAC identity, enabling seamless VM mobility. When a workload moves racks, it never needs to update its ARP table; the network automatically routes to the nearest local switch port.
"The ratio of server bandwidth to fabric bandwidth; modern AI clusters aim for 1:1 oversubscription for total performance."
"A VXLAN Tunnel Endpoint (VTEP) is the logical entrance/exit to the virtual fabric, typically mapped to the switch's Loopback IP."
"Utilizing passive optical taps allows for 100% visibility into fabric traffic without impacting the production packet forwarding path."
