Data Center Design
What the exam tests
Data Center Design tests spine-leaf topology, Layer 3 underlay, EVPN-VXLAN overlay, multihoming, oversubscription, automation, and failure-domain planning with Juniper QFX and Apstra.
Core rules
- Build a non-blocking or predictable oversubscription spine-leaf fabric; keep leaf-to-spine links equal cost.
- Use an L3 underlay (OSPF, IS-IS, or BGP) and EVPN-VXLAN for Layer 2 extension; avoid large spanning-tree domains.
- Use ESI-LAG or MC-LAG for server multihoming and anycast gateway for seamless VLANs.
- Place workloads by failure domain, not just rack density; separate management, storage, and compute traffic where needed.
- Automate fabric intent and validate with telemetry; document cabling and ASN/IP plans.
Worked example
A 20-rack general-compute pod uses 25 Gbps server ports. Each leaf has 48 x 25 Gbps downlinks = 1,200 Gbps. Four 100 Gbps spine uplinks per leaf give 400 Gbps, a 3:1 oversubscription. If the workload is latency-sensitive, use eight 100 Gbps uplinks for 1.5:1. With four spines, each leaf connects once to each spine; a single spine failure leaves 75% capacity, which meets N+1 for this pod. Deploy EVPN-VXLAN with two route reflectors, ESI-LAG to dual-homed servers, and anycast gateway on leaves. The reasoning is that oversubscription is a policy choice tied to traffic profile, not a universal “must be 1:1” rule.
Traps
Running spanning tree across the fabric; confusing overlay and underlay troubleshooting; using a single MLAG pair as the only redundancy; ignoring power and cooling per rack; and designing without an IP/ASN allocation plan.