Ceph: too many PGs per OSD

问题:

I configured Ceph with the recommended values (using a formula from the docs). I have 3 OSDs, and my config (which I've put on the monitor node and all 3 OSDs) includes this:

osd pool default size = 2
osd pool default min size = 1
osd pool default pg num = 150
osd pool default pgp num = 150

When I run ceph status I get:

 health HEALTH_WARN
        too many PGs per OSD (1042 > max 300)

This is confusing for two reasons. First, because the recommended formula did not satisfy Ceph. Second, and most puzzling, is that it says I have 1042 PGs per OSD, when my configuration says 150.

What am I doing wrong?

回答1:

Before setting PG count you need to know 3 things.

1. Number of OSD

ceph osd ls

Sample Output:
 0
 1
 2

 Here Total number of osd is three.

2. Number of Pools

ceph osd pool ls or rados lspools

Sample Output:
  rbd
  images
  vms
  volumes
  backups

Here Total number of pool is five.

3. Replication Count

ceph osd dump | grep repli

Sample Output:
 pool 0 'rbd' replicated size 2 min_size 2 crush_ruleset 0 object_hash rjenkins pg_num 64 pgp_num 64 last_change 38 flags hashpspool stripe_width 0
 pool 1 'images' replicated size 2 min_size 2 crush_ruleset 1 object_hash rjenkins pg_num 30 pgp_num 30 last_change 40 flags hashpspool stripe_width 0
 pool 2 'vms' replicated size 2 min_size 2 crush_ruleset 1 object_hash rjenkins pg_num 30 pgp_num 30 last_change 42 flags hashpspool stripe_width 0
 pool 3 'volumes' replicated size 2 min_size 2 crush_ruleset 1 object_hash rjenkins pg_num 30 pgp_num 30 last_change 36 flags hashpspool stripe_width 0
 pool 4 'backups' replicated size 2 min_size 2 crush_ruleset 1 object_hash rjenkins pg_num 30 pgp_num 30 last_change 44 flags hashpspool stripe_width 0

You can see each pool has replication count two.

Now Let get into calculation

Calculations:

Total PGs Calculation:

Total PGs = (Total_number_of_OSD * 100) / max_replication_count

This result must be rounded up to the nearest power of 2.

Example:

No of OSD: 3
No of Replication Count: 2

Total PGs = (3 * 100) / 2 = 150. Nearest Power of 150 to 2 is 256.

So Maximum Recommended PGs is 256

You can set PG for every Pool

Total PGs per pool Calculation:

Total PGs = ((Total_number_of_OSD * 100) / max_replication_count) / pool count

This result must be rounded up to the nearest power of 2.

Example:

No of OSD: 3
No of Replication Count: 2
No of pools: 5

Total PGs = ((3 * 100) / 2 ) / 5 = 150 / 5 = 30 . Nearest Power of 30 to 2 is 32.

So Total No of PGs per pool is 32.

Power of 2 Table:

2^0     1
2^1     2
2^2     4
2^3     8
2^4     16
2^5     32
2^6     64
2^7     128
2^8     256
2^9     512
2^10    1024

Useful Commands

ceph osd pool create <pool-name> <pg-number> <pgp-number> - To create a new pool

ceph osd pool get <pool-name> pg_num - To get number of PG in a pool

ceph osd pool get <pool-name> pgp_num - To get number of PGP in a pool

ceph osd pool set <pool-name> pg_num <number> - To increase number of PG in a pool

ceph osd pool set <pool-name> pgp_num <number> - To increase number of PGP in a pool

*usually pg and pgp number is same

回答2:

How I fixed it in 12.2.4 luminous:

Too many PGs per OSD (380 > max 200) may lead you to many blocking requests.

First you need to set:

[global]

mon_max_pg_per_osd = 800  # < depends on you amount of PGs
osd max pg per osd hard ratio = 10 # < default is 2, try to set at least 5. It will be
mon allow pool delete = true # without it you can't remove a pool 

Then restart all MONs and OSDs, one by one.

Check the value:

ceph --admin-daemon /var/run/ceph/ceph-mon.ceph2.asok config get  mon_max_pg_per_osd
ceph --admin-daemon /var/run/ceph/ceph-osd.3.asok config get osd_max_pg_per_osd_hard_ratio

Now look here:

rados lspools
ceph osd pool get .users.email pg_num

In my case by default pg_num was 128 or something like that (my cluster is 4 years old, it was a lot of upgrades a lot of changes). You can reduce it like that.

Be careful:

ceph osd pool create .users.email.new 8
rados cppool .users.email default.rgw.lc.new
ceph osd pool delete .users.email .users.email --yes-i-really-really-mean-it
ceph osd pool rename .users.email.new .users.email
ceph osd pool application enable .users.email rgw

If it wasn't enough, try to find another pool you can cut.