Data Center Networking Fundamentals and Architectures
Domain 1 tests how you turn workload requirements into a physical and logical data center topology: fabric shape, capacity, redundancy, and failure domains. Expect items that give port counts and ask for oversubscription, or describe a traffic pattern and ask which topology fits.
Fabric topology
- Modern designs use a two-tier spine-leaf Clos: every leaf connects to every spine, so any host-to-host path is leaf-spine-leaf with consistent latency.
- East-west traffic dominates; the fabric is Layer 3 with ECMP instead of one large Layer 2 domain.
- A three-tier access/aggregation/core design survives mainly for legacy or small sites; it adds hops and spanning-tree risk.
Capacity and redundancy
- Oversubscription equals downstream host-facing bandwidth divided by upstream spine-facing bandwidth. 1:1 is non-blocking; 3:1 is common for general compute.
- Count redundancy per failure domain: dual power feeds, dual supervisors, uplinks to two different spines, and N+1 or N+N power and cooling.
- Jumbo frames (9000 MTU) help storage and backup traffic but must be consistent end to end, including storage arrays and hypervisors.
Worked example
A leaf has 48 host ports at 25 Gb/s and 8 uplinks at 100 Gb/s. Downstream is 48 x 25 = 1200 Gb/s. Upstream is 8 x 100 = 800 Gb/s. Oversubscription is 1200/800 = 1.5:1. If the design target is 3:1, the leaf is over-provisioned northbound and could drop to 4 uplinks (400 Gb/s, giving 3:1). The ratio is directional and must be computed from ports actually cabled and used, not from chassis maximums.
Traps that cost marks
- Computing oversubscription with every port populated rather than design-populated ports.
- Assuming a Layer 2 fabric scales like a Layer 3 Clos; broadcast and unknown unicast do not follow ECMP.
- Ignoring optics reach and FEC requirements when mixing 25G and 100G links.
- Treating a dual-homed host as fully redundant when both links land on the same leaf.
Exam framing
Questions often hide the real constraint in the wording, such as a rack power limit that caps port count, or a latency-sensitive workload that rules out extra tiers. Read for the constraint before calculating.