Ask a room of architects which distributed switch model they're using and half of them will answer with a number of switches. Ask how many physical NICs are in each host and you'll watch the answer change.

The model is the first one, but the NIC count is what actually decides it.

The Models on the Menu

Every host in a VCF deployment passes briefly through a standard switch and gets moved to a distributed switch during the Installer's run, so the layout has to be decided before bring-up. In my own practice the question comes up early, for the management domain and again for each workload domain, which is where the choice gets revisited.

A single vDS is the default. Management, vMotion, vSAN, and the NSX overlay all share one switch and one set of uplinks. Storage Separation moves storage traffic onto a second vDS: management, vMotion, and NSX stay on the first, and the second carries vSAN or other storage. Workload Separation does the same for NSX, giving the overlay its own switch. A vSAN storage cluster gets a variant of its own, with a dedicated vDS for its vSAN traffic and a recommendation to turn Network I/O Control off on it after deployment. The design library also describes Storage plus Workload Separation, and a Dedicated Edge Cluster pattern that builds on Workload Separation.

Whether a domain can move from one model to another after bring-up is something I'd confirm in the documentation for your target version. I wouldn't plan around it.

A Second Switch Needs a Second Set of NICs

This is where the diagram stops matching the budget. The installer's two-switch profiles create the second switch with its own physical NICs, and the design library gives each separation model a NIC count of its own. Storage Separation and Workload Separation are four-NIC designs. The combined Storage and Workload Separation model needs six for redundancy.

Two NICs buy you one switch with redundancy, or two switches without it.

That's the real constraint on a two-NIC host. Either the second switch can't be built, or each switch gets one NIC and loses its failover path. Neither is a choice anyone makes on purpose, but both are what a design document ends up describing when the hardware order was placed first.

The vSAN guidance lines up with this. It recommends two NICs per host for small-scale workloads and four for large or I/O-intensive ones. Those are recommendations, not hard minimums, but I wouldn't want to argue a two-NIC build into a separated design in front of a customer.

It's worth saying out loud in the first design session, not the fourth. Hardware gets quoted early and changes slowly.

Teaming, MTU, and the Physical Side

The physical network gets a vote too. The recommended pattern is to connect each host to two or more top-of-rack switches, and each connection between a fabric switch and an ESX host should work as an independent uplink. Static link aggregation and EtherChannel aren't supported. LACP to the host is supported but not recommended, and it needs LACP fallback mode on the first physical NIC. Aggregation between the fabric switches themselves is fine. If the network team's standard is LACP on every server port, that conversation needs to happen before bring-up.

The teaming defaults are worth knowing because they encode a reason. Management and vMotion port groups use "Route based on physical NIC load." Storage port groups default to physical NIC load too. In the models that give storage its own switch, the design library recommends changing storage to Failover Order during deployment, which keeps each storage flow on a single top-of-rack switch so it doesn't cross the inter-switch link. VM-management port groups default to ephemeral port binding and the rest to static. Network I/O Control is enabled by default, with shares of Normal for management, Low for vMotion, and High for both VM and storage traffic.

Jumbo frames need to be configured across the whole path: VMkernel ports, virtual switches, physical switches, and routers. A setting of 9000 is the recommendation for the vSAN, vMotion, and overlay paths. It isn't a blanket requirement, and management VMkernel ports are recommended at 1500. The firm number is the overlay's: VCF requires 1700 on the TEPs, though Geneve itself needs only 1600. Geneve packets carry the don't-fragment flag, so a hop that can't carry them doesn't slow down gracefully. It drops them. That's how you build something that pings fine and then falls over under real load.

One Switch, One Cluster, Mostly

The standing recommendation is not to share a vDS across different vSphere clusters. The documented reasons are independent lifecycle management and a smaller fault domain. That's still the recommendation, with a footnote: VCF 9.1.1 introduced support for sharing a vDS across multiple clusters in a domain. It's limited to greenfield deployments with Layer 2 clusters, and it isn't supported for Layer 3 clusters or for stretched ones.

So readers on 9.1.1 will see the option in the product. My advice is the same as the design library's: the tidy-looking single switch makes a change on one cluster a change on all of them, and that connects back to the domain-versus-cluster question from a few weeks ago. A boundary you don't draw in the switch layer tends to show up in the change window. Before 9.1.1, the option isn't there, so the question doesn't arise.

Questions I Ask Before Anyone Picks a Switch Model

How many physical NICs does each host have today, and is that number still changeable?

Does the storage design need traffic isolated from management, or is shared bandwidth with sensible NIOC shares good enough?

Does the physical network team expect LACP, and who has told them it's supported but not recommended here?

Is the MTU confirmed along the whole path for the traffic types that need it, not only on the host-facing ports?

If any cluster shares a vDS with another on 9.1.1, who signed up for the shared change window?

If you've got a switch diagram and a purchase order that disagree, send me a DM. It's a quick comparison and a much cheaper one than finding out in week six.