Traditional virtualization separates servers (compute) from a storage array (SAN or NAS). Hyperconverged infrastructure (HCI) puts both in every node: each server runs VMs and contributes its local disks to a distributed storage pool. You scale by adding identical nodes.
Before you start
- Background: Ceph for virtualization and Proxmox clusters.
Traditional vs hyperconverged
| Three-tier (servers + SAN) | Hyperconverged | |
|---|---|---|
| Storage location | Dedicated array | Local disks in every node |
| Scaling | Scale compute and storage separately | Add nodes, both grow together |
| Single points of failure | Array controllers (usually redundant) | Spread across nodes |
| Network | Storage fabric (FC, iSCSI) | Fast Ethernet between nodes is critical |
| Management | Several tools | Often one interface |
Examples
- Proxmox VE with built-in Ceph.
- VMware vSAN.
- Microsoft Storage Spaces Direct with Hyper-V.
- Commercial HCI appliances from several vendors.
Benefits
- Fewer device types and a simpler rack.
- Live migration and HA without a separate SAN.
- Incremental growth: buy another node when needed.
Drawbacks
- Storage replication uses CPU, RAM and network on the same nodes that run VMs.
- Compute and storage scale together, which can waste one if you only need the other.
- A minimum of three nodes is typical for safe quorum and replicas.
- Troubleshooting distributed storage needs skill.
Sizing tips
- Reserve CPU and RAM per node for the storage software (for Ceph, memory per OSD).
- Use at least 10 GbE, ideally 25 GbE, with redundant links for storage traffic.
- Plan capacity so the cluster can lose a node and still re-replicate data.
- Use enterprise SSDs/NVMe with power-loss protection.
Frequently asked questions
Is HCI good for small businesses?
A three-node HCI cluster can replace servers plus a SAN, but a single well-backed-up host is simpler if downtime for repairs is acceptable.
Can I mix node types?
Possible, but similar nodes keep performance and data placement balanced.
Do VPS providers use HCI?
Some do; others prefer local NVMe per host for performance and simplicity.