Ceph planner for Proxmox VE
Enter nodes, OSDs and pools. The planner shows usable capacity, placement groups per pool, what happens when a node or an OSD fails, and the pveceph commands for exactly this layout. Everything is calculated in your browser; nothing is sent anywhere.
Start from an example
- Usable
- 3.84 TB
- Planned (85 %)
- 3.26 TB
- PGs per OSD
- 128
Your cluster
Cluster
Switch off to edit every node, add OSD groups with other sizes or device classes.
Redundancy and pools
Sum of all shares: 100 %
Failure domain
CRUSH never puts two copies or chunks of a PG into the same failure domain.
Network
Ceph needs at least 10 Gbit/s. A separate cluster network carries replication and recovery.
Planning
Used for the recovery estimate of the failure simulation.
Result
Usable 3.84 TB, planned 3.26 TB, 128 PGs per OSD, 3 notes.
- Raw
- 11.52 TB
- 6 OSDs on 3 nodes
- Usable
- 3.84 TB
- after replication and EC
- Planned (85 %)
- 3.26 TB
- 128
- target 100
- MON / MGR
- 3 + 3
- 3 monitors and 3 managers
Capacity
- Usable after redundancy / Planned, up to nearfull: 3.26 TB
- Redundancy overhead
- Reserve until full
- Writes blocked
- · backfillfull 90 % · full 95 %
Raw capacity 11.52 TB, of which 3.84 TB usable after redundancy. Plan with 3.26 TB: Ceph warns at 85 %, stops backfilling at 90 % and blocks writes at 95 %.
Cluster layout
Select nodes and OSDs to simulate their failure; select one again to undo. Any number can fail at once. Block height is proportional to the OSD size.
- MONMGR
- MONMGR
- MONMGR
Legend
- NVMe
- Simulated failure
- Recommended monitor (MON) and manager (MGR) nodes
Failure simulation
No failure simulated. Select nodes or OSDs in the cluster layout; any number can fail at once.
| Pool | Share | size / k+m | Raw value | pg_num | PGs per OSD | target_size_ratio |
|---|---|---|---|---|---|---|
| vmstore | 100 % | 3 | 200 | 256 | 128 | 1 |
| Total | 128 | |||||
Alternative: enable the autoscaler (pg_autoscale_mode on) with the target_size_ratio shown. It aims at about 100 PGs per OSD; the static pg_num is the starting point.
Warnings and notes
Note: No self-healing after a host failure
Pool vmstore
Note: No separate cluster network
Note: The network limits the NVMe OSDs
Commands
These commands run as root and are a generated suggestion, not a reviewed configuration for your cluster. Read every line before running it, and try it in a test environment first. Details and liability
# ---------------------------------------------------------------------
# Generated suggestion - NOT a reviewed configuration for your cluster.
# It was produced from the values entered in a planning tool that has
# no access to this cluster and cannot check its state, its hardware or
# its data. Placeholders must be adapted. Capacity and recovery figures
# are estimates based on documented assumptions.
# Read every line before running it as root, and try it on a test
# system first.
# ---------------------------------------------------------------------
# Ceph setup for Proxmox VE 9 - generated plan, review before running.
# Planned layout:
# node1: 2 x 1.92 TB nvme
# node2: 2 x 1.92 TB nvme
# node3: 2 x 1.92 TB nvme
# 1) On every node (or through the GUI): install Ceph
# Use --repository enterprise when you have a subscription.
pveceph install --repository no-subscription
# 2) On the first node only: initialise Ceph.
# Replace 10.10.10.0/24 with your Ceph public network.
pveceph init --network 10.10.10.0/24
# 3) Monitors and managers - run each line on node1, node2, node3
pveceph mon create
pveceph mgr create
# 4) OSDs - run on each node, once per device.
# Replace sdX with the real block device name from lsblk.
# Add --db_dev /dev/nvme0n1 to put the RocksDB of an HDD OSD on flash.
pveceph osd create /dev/sdX
# 5) Pools
pveceph pool create vmstore --size 3 --min_size 2 --pg_num 256 --pg_autoscale_mode on --add_storagesReview before running: device names and networks are placeholders.
Every formula, constant and assumption behind these numbers: read the methodology
Something wrong, unclear or missing? Send feedback by e-mail
Frequently asked questions
Why 512 and not 400 PGs?
400 is the raw pgcalc value before rounding; pg_num must always be a power of two. 512 is the power of two nearest to 400 (256 would be 36 % below it), so no further guard is needed — the result rounds up to 512 directly.
The rule, including the case where the nearest power of two is more than 25 % below the raw value and the next one up is used instead, is explained in full on the placement groups article.
Do three nodes suffice?
Three nodes is the minimum for a highly available Ceph cluster and is enough to place a replicated size 3 pool at all. It is not enough for Ceph to self-heal after a node fails: with exactly three hosts and size 3, every host already holds a copy, so there is nowhere to rebuild the missing one until the failed node returns.
The planner reports this as an informational finding (hosts == size) rather than an error, because the cluster is still valid and data stays available. A fourth node turns it into a cluster that heals itself.
TB or TiB?
The planner defaults every size input to decimal terabytes (TB, 10¹² bytes), because that is how drives are sold and labelled. Ceph itself reports capacity in binary tebibytes (TiB, 2⁴⁰ bytes), which is about 9 % larger per unit than a TB.
A unit toggle switches the display between the two; internally every computation happens in bytes, so switching the display unit never changes a result, only how it is written.
Is size 2 okay?
size 2 is placeable and halves the raw capacity needed for the same usable capacity compared with size 3. The trade-off: it survives no double failure at all. With the recommended min_size 2, a single OSD failure already blocks IO on the affected placement groups, because only one copy then remains.
The planner always marks size 2 as a warning. It can be a reasonable, deliberate choice for data that is easy to reproduce or genuinely not critical — it is not a reasonable default.
Why plan with 85 % instead of the full usable capacity?
0.85 is Ceph’s own mon_osd_nearfull_ratio: the fill level at which Ceph starts raising a health warning. Filling a cluster to its raw usable capacity means it is already past that point, with no room left to absorb even one failure without also tripping backfillfull (0.90) on the way to recovery.
The planner therefore shows "planned" capacity as usable capacity × 0.85 next to the raw usable figure, so the number actually worth planning against is visible on its own.
Can erasure coding be used for VM disks?
Technically yes — Proxmox supports erasure coded pools with application rbd like any other. It costs meaningfully more write amplification and CPU per write than replication, because every write recomputes coding chunks instead of duplicating bytes.
Common practice, not a hard rule enforced anywhere, is to reserve erasure coding for backup, archive and CephFS data, and to keep latency-sensitive VM storage on a replicated pool. The planner surfaces this as an informational note rather than a warning, precisely because it is a judgement call.
Autoscaler or a fixed pg_num?
The planner always computes both: a fixed pg_num from the pgcalc formula, and the autoscaler alternative (pg_autoscale_mode on with target_size_ratio). Proxmox’s own default for a new pool is 128 PGs with the autoscaler left in warn mode — it suggests a resize but does not perform one automatically.
A reasonable approach is to start from the computed fixed value and switch the autoscaler on deliberately, rather than leaving it in warn mode indefinitely.
How accurate is the recovery time estimate?
Deliberately not very — it is a rough range, not a prediction. It assumes half the network link and 30 % of surviving OSD write throughput are available for recovery traffic, using conservative, named throughput figures per device class, and shows the result as a 0.5× to 2× range around that point estimate.
Real recovery time depends on osd_recovery tuning, CPU, concurrent client load and how evenly data is spread — none of which the planner has any way to know. Every constant behind the estimate is named on the methodology page.
What happens to the data I enter?
Nothing leaves your browser. Every computation — capacity, PGs, findings, recovery estimate, generated commands — runs client-side; the planner has no backend and performs no network requests at runtime.
The permalink that shares a setup contains the plan itself, compressed into the URL — there is no server-side storage to share from. If the site is hosted behind a web server, that server’s own access logs (if any) are controlled by its operator, not by this planner.
Are the generated commands safe to paste into a shell?
Review them before running them, as with any command a tool hands you for a root shell. That said, the planner treats generated commands as something a human will actually run: node and pool names are restricted to a safe, narrow pattern before they are ever interpolated into a command, and the network addresses shown are always the same fixed placeholders, never a value derived from what was typed.
No commands at all are generated for a layout the planner has marked as not placeable — an error-severity finding suppresses the whole command output, on both tabs, so an unsafe or invalid setup never hands out copy-ready commands to create it.