Thanks,
I found this thread [1] recommending offline compaction with large
OMAP/META data on the OSDs. We'll try that first and see if it helps.
We're still missing some details about the degredation, I'll update
this thread when we know more.
The PGs are balanced quite perfectly and we don't have spillover.
[1]
https://lists.ceph.io/hyperkitty/list/ceph-users@xxxxxxx/thread/EDL7U5EWFHSFK5IIBRBNAIXX7IFWR5QK/
Zitat von "Szabo, Istvan (Agoda)" <Istvan.Szabo@xxxxxxxxx>:
Hi,
Last year this caused my long lasting issue, the autoscaler
suggestion was completely irrelevant about pg number.
I had an objectstore cluster with 2B objects, most of them small
like 50kb, and the only solution to avoid the slow ops in my cluster
was to:
- increase the pg number
- Use the balancer with max deviation 1
- Remove wal+rocksdb from a separated device because it was
continuously spill overed
- 4osd for each 15TB ssds
Now the cluster is kind of stable in the last half year.
Istvan Szabo
Senior Infrastructure Engineer
---------------------------------------------------
Agoda Services Co., Ltd.
e: istvan.szabo@xxxxxxxxx
---------------------------------------------------
-----Original Message-----
From: Eugen Block <eblock@xxxxxx>
Sent: Tuesday, July 26, 2022 4:39 PM
To: ceph-users@xxxxxxx
Subject: Impact of many objects per PG
Email received from the internet. If in doubt, don't click any link
nor open any attachment !
________________________________
Hi *,
are there any known limitations or impacts of (too) many objects per
PG? We're dealing with a performance decrease on Nautilus (I know,
but it can't be upgraded at this time) while pushing a million
emails (many small objects) into the cluster. At some point, maybe
between
600.000 and 900.000 emails or so, the client requests became slower
(CephFS kernel clients) although all entities look fine, the MDS
daemons are not overloaded, the OSDs not fully utilized (HDDs with
shared rocksDB on SSDs). The HDD OSDs currently have around 100 PGs
per OSD with 45 GB per PG (quite a lot) and around 180.000 objects
per PG. The main data pool has an EC profile with k=4 m=5.
We want to increase the pg_num anyway, we expect a general
performance increase after the pg splitting, but we're still
wondering where the limits are. For example, if we increase
mon_max_pg_per_osd to more than 250 we could split the PGs even more
to reduce PG size and number of objects per PG, if the OSDs can cope
with it utilization-wise, but I'm not sure if that's a good idea.
Are there any other tunables we could tweak instead to reduce the
performance impact?
Any comments or references are highly appreciated!
Thanks,
Eugen
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an
email to ceph-users-leave@xxxxxxx
________________________________
This message is confidential and is for the sole use of the intended
recipient(s). It may also be privileged or otherwise protected by
copyright or other legal rules. If you have received it by mistake
please let us know by reply email and delete it from your system. It
is prohibited to copy this message or disclose its content to
anyone. Any confidentiality or privilege is not waived or lost by
any mistaken delivery or unauthorized disclosure of the message. All
messages sent to and from Agoda may be monitored to ensure
compliance with company policies, to protect the company's interests
and to remove potential malware. Electronic messages may be
intercepted, amended, lost or deleted, or contain viruses.
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
