Ceph PG calculator

How pg_num is calculated

The rule is the one PGCalc uses, described in the Ceph documentation: start from the target PGs per OSD, multiply by the OSDs of the class and by the pool's share of the data, and divide by the width of the pool, that is, the number of copies with replication or k+m with erasure coding.

x = target × OSDs × share / width
x = max(x, OSDs / width)
pg_num = nearest power of 2 to x
if pg_num < 0.75 × x then pg_num = 2 × pg_num

Worked example: 12 OSDs, a single 3-way replicated pool, target 100. x is 100 × 12 / 3 = 400; the nearest power of 2 is 512, which is not below 75% of 400, so pg_num is 512. The rule doubles the result when the nearest power of 2 is more than 25% below x: x = 1,365 gives 1,024, x = 1,366 gives 2,048.

The starting target is 100, the default of mon_target_pg_per_osd. The PGCalc rule uses 200 when the OSD count may double, so that you do not have to raise the pg_num of the pools soon after creating them.

PGCalc or the autoscaler

With pg_autoscale_mode on, Ceph adjusts pg_num on its own, based on the data stored and the data expected. In autoscaler mode the calculator sets target_size_ratio for each pool to its share within the class: it is a relative weight, not a hard quota, and the autoscaler normalises it across the pools that share the same OSDs. Pools that will hold most of the data get bulk, which makes them start with many PGs instead of growing a little at a time.

In PGCalc mode the calculator sets the autoscaler to warn: pg_num stays as calculated, and Ceph warns if it considers it far from optimal.

PG instances per OSD: average and maximum

The average is not enough. In the model every pool spreads across the nodes in proportion to capacity, with at most one instance of each PG per node, and the instances of all pools add up on the same OSDs. Ceph applies mon_max_pg_per_osd (250 by default in Squid and Tentacle) to the cluster average: above it, Ceph raises TOO_MANY_PGS and refuses pools or pg_num values that would exceed it. The calculator also compares it with the estimated maximum on a single OSD, as a conservative check, and shows the hard limit of a single OSD, three times as much with osd_max_pg_per_osd_hard_ratio set to 3: above it the OSD creates no new PGs. It repeats the check after losing a node, when the instances of the lost node are redistributed over the others.

Worked example: three nodes with 12, 1 and 1 OSDs, one 3-way replicated pool with 512 PGs, host failure domain. The average is about 110 instances per OSD, but every PG has a copy on each node: the OSDs of the two single-drive nodes carry 512 instances each, more than twice mon_max_pg_per_osd.

Ready-made commands

The commands come from the same pool table. With ceph the calculator generates: a replicated CRUSH rule per device class; pool creation with size and min_size; for erasure coding, a profile with crush-device-class and crush-failure-domain; allow_ec_overwrites for RBD and CephFS; the application enabled on the pool and the autoscaler mode.

With pveceph pool create and --erasure-coding, Proxmox VE creates two pools: <name>-data with erasure coding, the given pg_num and allow_ec_overwrites already enabled, and <name>-metadata, replicated with 32 PGs. --size, --min_size and --crush_rule apply to the metadata pool, so the min_size of the data pool is set with a separate command. For EC pools not used for RBD the calculator adds --add_storages 0, because with erasure coding pveceph also creates an RBD storage entry in Proxmox VE by default. On Proxmox VE, CephFS is created with pveceph fs create, and an EC pool is added as a data pool with ceph fs add_data_pool.

Formulas and sources

Model 1.2.0, dated 19 October 2026, shared with the Ceph calculator. The behaviour of pveceph with erasure coding was verified in the Proxmox VE source code, not only in the documentation.

• Ceph, placement groups and the autoscaler
• Ceph, the PGCalc rule
• Ceph, pools, size and min_size
• Ceph, erasure coding
• Proxmox VE, Ceph pools and erasure coding
• Proxmox VE, pveceph syntax
• Proxmox VE, pool creation source code