Networking

How the layers fit together, from physical NICs up to the networks a VM attaches to.

VS-HCI networking is software-defined. The physical network carries packets between servers; everything a virtual machine sees is built on top by the platform, consistently across every host.

The layers

   VM  ──►  network on a distributed switch
              │
              ▼
          distributed switch  (L2GDS / L3GDS / SGDS / RGDS / MGDS)
              │
              ▼
          uplink  (a bond of NICs behind a bridge, on each member host)
              │
              ▼
          physical NICs ──► your switches

Read it bottom-up:

  1. Uplinks claim physical NICs, bond them, and put a bridge on top. One uplink spans many hosts, with a member on each. See Uplinks and bonds.
  2. Distributed switches sit on an uplink and provide networks that exist identically on every member host. See Distributed switches.
  3. VMs attach to a network on a switch. A VM keeps its network when it moves hosts, because the switch is on all of them.

Why nothing is configured at install time

The installer configures exactly one thing: the management link. Every other network is created afterwards, from the console.

That is deliberate. An uplink picks NICs, bond mode, LACP parameters and MTU together and can verify them against your switches — decisions that need to be made as a set, by someone who can see the result. A network pre-created during installation would be a guess.

Encapsulated or not

The distinction that drives most of the design:

  Not encapsulated Encapsulated (Geneve)
Used by L2 switches, storage, replication, management L3 overlay fabrics
VLANs Handed off to your switches as 802.1Q Carried inside tunnels
Needs underlay IPs No Yes, one per host
MTU The uplink’s wire MTU Wire MTU minus 58 bytes
Your switches must Trunk the VLANs Route between hosts and pass UDP 6081

A VLAN-based network hands traffic to your existing switches and your existing router is the gateway. An overlay network carries tenant traffic inside tunnels between hosts, and the platform routes it.

MTU

Geneve encapsulation costs 58 bytes. The platform tracks two numbers per uplink:

  • Underlay MTU — what the physical path must carry.
  • Fabric MTU — what the overlay can offer, 58 bytes smaller.

Four profiles are selectable:

Fabric MTU Underlay MTU  
9158 9216 Maximum jumbo
8942 9000 Default — standard jumbo
4442 4500  
1442 1500 Legacy, for a network that cannot do jumbo

The default assumes your switches carry 9000-byte frames. If they do not, choose 1442/1500 — and expect storage replication to be slower for it.

Note — Networks that are not encapsulated — storage, replication, management — carry the uplink’s wire MTU, not the overlay MTU. They also follow the uplink: raise the uplink to jumbo and those ports come with it.

The control plane stays out of the data path

The platform distributes topology to every host and then gets out of the way. Packets between two VMs go host-to-host; they do not traverse a controller. A control plane outage does not stop traffic that is already flowing.

Where to go next

Task Page
Claim NICs, build bonds Uplinks and bonds
Give VMs networks Distributed switches
Isolated routed domains Fabrics
Get traffic in and out Routing and NAT
Peer with your routers BGP
Filter traffic Security policies