A cluster with no switch: direct mesh, and what it changes in your rack
A 100 Gbit/s network card has become affordable. The switch that goes with it, far less so: a few fast switches can cost more than new servers. For a small cluster there is a way out: not buying any. Here is how, how far, and what it changes on your hosting bill.
Estimate my cost →The dilemma: the card is affordable, the switch is not #
When sizing a hyperconverged virtualization cluster, shared storage demands a lot of bandwidth between servers: every write is copied to two other nodes, and a rebuild after a failure saturates whatever you give it. You quickly want 25 or 100 Gbit/s. The cards exist at reasonable prices. Switches in that class remain a capital item in their own right, plus their maintenance, their power draw, their rack space and their replacement the day they fail.
What the nodes say to each other never leaves the rack #
This is the observation that unlocks everything. Two of a cluster's flows only ever address the other nodes: shared storage, and the heartbeat that constantly checks each one is alive. Neither needs to talk to the outside world. Only virtual machine traffic and administration must leave. In other words: the switch is only indispensable for the part that leaves the rack.
Direct mesh, in practice #
In a mesh, each server is cabled straight to each other one, with no equipment in between. For a three-node cluster: the outside link stays a star, with two bonded links to your usual switches; the heartbeat gets a four-port 1 Gbit/s card per server, two ports towards each of the other two; storage gets a four-port fast card per server, cabled the same way. The result is a very fast storage back end and a very stable heartbeat, for the price of three cards, with no fast switch to buy, power or replace.
Two separate meshes, never one #
The temptation is to run everything over the same mesh to save ports. That is the same mistake as putting the heartbeat in a VLAN of the storage network: the cable is shared, so the queue is too. When a rebuild saturates the link, the heartbeat waits its turn, overruns its deadline, and the cluster reboots servers that were perfectly fine. Two physically distinct meshes: a slow one dedicated to the heartbeat, a fast one for storage.
Where it stops: the port count #
The mesh has a simple arithmetic limit: each server needs two ports towards each other one. At three nodes that is four ports per zone per server. At five nodes, eight. And crucially, adding a server forces you to add two ports on every existing server, so to open them all. A mesh is therefore excellent from two to five nodes when growth is not planned, and poor as soon as it is. Beyond that, back to a star, and the switch becomes the right tool again.
The ring trap #
When card slots start running out, the idea comes naturally: rather than linking everyone to everyone, link each server to its two neighbours, in a loop. It works technically, and it is a bad idea. In a mesh, every node is one hop from every other. In a ring, data crosses intermediate servers: latency rises, and each server additionally carries traffic that is none of its business, which cuts its available bandwidth. Storage is never faster than its slowest link, and the heartbeat responds to unstable latency with reboots. If slots are short, reduce the number of links or go back to a star, but do not close the loop.
What it changes in your rack, and therefore in your quote #
This is where the subject stops being technical. Removing the storage switches removes rack units, removes their power draw from the capacity you subscribe, and removes the cross-connects that were dedicated to them. On a three-node cluster the saving lands in rack units, kVA and cross-connects, that is, in the three lines that make the price of hosting. Our Digital Architect prices the cluster and produces the matching requirement sheet, space, power and cooling, with the mesh option whenever the cluster size allows it.
FAQ #
No, and it even removes a point of failure: what does not exist cannot fail and does not need updating. The trade-off is not reliability, it is growth: adding a server forces you to open all the others to add ports.
Two to five in practice. At five nodes each server already carries eight ports per zone. Beyond that, the number of cards and cables outweighs the saving on switches.
No. A shared cable means a shared queue: a storage rebuild delays the heartbeat, and a late heartbeat reboots healthy servers. Two physically distinct meshes.
It depends on the switches removed, but the saving reads on three lines: the rack units they occupied, their share of the subscribed power, and the cross-connects dedicated to them.
Written on 12 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers