Proxmox VE Hardware Sizing Calculator (3+ Nodes)

Estimate per-node CPU, memory, NVMe and networking for a Proxmox VE cluster from three nodes up, compare node counts, and review Ceph must-haves and what to check when adding nodes.

Enter your VM count, average size and number of nodes to estimate the CPU, memory, disks and networking each Proxmox VE node needs, from three nodes up as you add more. The sizing lets the remaining nodes carry every VM if one node fails, and includes the MOAVI management VM.

Adding nodes to a three-node cluster

  • If the three nodes use a full mesh, move Ceph to two 25 GbE switches before adding a fourth node. The mesh is meant for three nodes.
  • Give the new node the same number and size of OSD disks as the others.
  • New OSDs trigger rebalancing, which adds network, disk and memory load. Do it outside business hours.
  • When mixing CPU generations, set VMs to a common CPU type (such as x86-64-v2/v3) so live migration keeps working.
  • Join the new node with a dedicated corosync link plus a second link on another physical network.
  • Bring the new node to the same Proxmox VE and Ceph versions before it joins.

Must-haves for Ceph

  • Use NVMe SSDs, enterprise grade with power-loss protection (PLP) only. No consumer SSDs.
  • Give Ceph its own low-latency network: 25 GbE or faster for NVMe (10 GbE at minimum), with redundant ports and switches.
  • Three nodes can be cabled directly to each other (full mesh) without a 25 GbE switch. Use switches beyond three nodes or if you plan to grow.
  • Keep cluster communication (corosync) on its own 1 GbE link. Sharing it with Ceph traffic can cost the cluster its quorum.
  • Use an HBA or direct NVMe instead of a RAID controller. Ceph must manage the disks itself.
  • Give every node the same number and size of OSD disks, at least 12 in the cluster (four or more per node).
  • Allow at least 8 GB of memory and one CPU core (thread) per OSD. Fast NVMe OSDs can use four to six threads.
  • Run with size 3 and min_size 2 (never 1), and keep usage well below the 85% nearfull warning.

Recommended for every cluster

  • Use three identical servers where possible.
  • Use 64-bit server CPUs with Intel VT-x or AMD-V.
  • Separate management, VM and storage networks, with redundant NICs.
  • Use ECC memory and redundant power supplies.
  • Mirror two PLP SSDs for the boot disk.
  • With external SAN storage, use redundant paths (multipath).

How we calculate

  • The remaining nodes can run every VM if one node fails (N+1).
  • CPU uses the overcommit ratio you enter; memory is not overcommitted.
  • Each node reserves 2 CPU cores and 4 GB of memory for Proxmox VE itself.
  • Ceph gets 2 cores and 4 GB for MON and MGR, plus 2 cores (4 threads) and 8 GB per NVMe OSD.
  • Ceph targets 70% usage on three nodes (room for one failed OSD). From four nodes, it targets 63%, 68% and then 70% so re-replicating one failed node still stays below the 85% warning.
  • A three-node Ceph cluster runs with two copies until a failed node is replaced, so plan for fast replacement.
  • ZFS replication copies each VM to one other node, on mirrors kept below 80%, with 1 GB of memory per TB used.
  • NVMe sizes are picked from 1.92, 3.84, 7.68 and 15.36 TB so each node holds four to eight.
  • The MOAVI management VM is included: 4 vCPU, 8 GB and 100 GB up to 1,000 VMs, otherwise 8 vCPU, 16 GB and 200 GB.
  • Results are estimates; the actual build is confirmed in an assessment.