tl;dr: this cluster is up again, thank you all (Mark, Wout, Paul
Emmerich off-list)!
First we tried to lower max- and min_pg_log_entries on a single running
OSD, without and with restarting it. There was no effect. Maybe because
of the unclean state of the cluster.
Then we tried ceph-objectstore-tool trim-pg-log on an offline OSD. This
has to be called per PG that is stored on the OSD. At first it seemed to
be much too slow, took around 20 minutes. But the following PGs were
much faster (like 1 minute). The trim part of the command was always
fast, but the compaction part took a long time the first time.
CEPH_ARGS="--osd-min-pg-log-entries=1500 --osd-max-pg-log-entries=1500"
ceph-objectstore-tool --data-path /var/lib/ceph/osd/ceph-$OSD --pgid $pg
--op trim-pg-log
Thanks to these pglog trimmings, memory consumption was reduced and we
could bring up all OSDs. Then recovery was quite fast, no big
backfilling. We checked a bit later in the evening, there was plenty of
free RAM.
Next morning, again free memory was very tight. Although it looked
differently. dump mempools showed buffer_anon as biggest (should this
not be tuned down by "osd memory target"?). But also osd_pglog (although
all PGs were active+clean?).
Soon there was another OOM killer. Again, we treated this OSD with
trim-pg-log to bring it back.
Then we decided to try again to reduce the pg_log parameters, cluster
wide (from default 3000 to 2000). This time it worked, memory was
released :-)
Then we added some RAM to get more to the safe side.
Some more background. As already mentioned, the number of PGs per OSD is
ok, but there is a lot of small objects (nearly 1 billion), mostly S3,
in an EC pool 8+3. So the number of the objects lieing on the OSDs
(chunks? shards?) is about 10 billions in total. Per OSD (510 of type
hdd) this is probably quite a lot. Maybe also a reason for high pglog
demand. And it is not equally distributed, HDDs are 4 TB and 8 TB.
Another point: the DB devices lie on SSDs. But they are too small
nowadays, the sizing was done years ago, for Filestore.
Last not least, probably the trigger was a broken HDD on the Sunday
before. Rebalancing then takes several days and was ongoing when the
problems started.
Cheers
Harry
On 14.05.20 11:08, Wout van Heeswijk wrote:
Hi Harald,
Your cluster has a lot of objects per osd/pg and the pg logs will grow
fast and large because of this. The pg_logs will keep growing as long as
you're clusters pgs are not active+clean. This means you are now in a
loop where you cannot get stable running OSDs because the pg_logs take
too much memory, and therefore the OSDs cannot purge the pg_logs...
I suggest you lower the values for both the osd_min_pg_log_entries and
the osd_max_pg_log_entries. Lowering these values will cause Ceph to go
into backfilling much earlier, but the memory usage of the OSDs will go
down significantly enabling them to run stable. The default is 3000 for
both of these values.
You can lower them to 500 by executing:
ceph config set osd osd_min_pg_log_entries 500
ceph config set osd osd_max_pg_log_entries 500
When you lower these values, you will get more backfilling instead of
recoveries but I think it will help you get through this situation.
kind regards,
Wout
42on
On 13-05-2020 07:27, Harald Staub wrote:
Hi Mark
Thank you for your feedback!
The maximum number of PGs per OSD is only 123. But we have PGs with a
lot of objects. For RGW, there is an EC pool 8+3 with 1024 PGs with
900M objects, maybe this is the problematic part. The OSDs are 510
hdd, 32 ssd.
Not sure, do you suggest to use something like
ceph-objectstore-tool --op trim-pg-log ?
When done correctly, would the risk be a lot of backfilling? Or also
data loss?
Also, to get up the cluster is one thing, to keep it running seems to
be a real challenge right now (OOM killer) ...
Cheers
Harry
On 13.05.20 07:10, Mark Nelson wrote:
Hi Herald,
Changing the bluestore cache settings will have no effect at all on
pglog memory consumption. You can try either reducing the number of
PGs (you might want to check and see how many PGs you have and
specifically how many PGs on that OSD), or decrease the number of
pglog entries per PG. Keep in mind that fewer PG log entries may
impact recovery. FWIW, 8.5GB of memory usage for pglog implies that
you have a lot of PGs per OSD, so that's probably the first place to
look.
Good luck!
Mark
On 5/12/20 5:10 PM, Harald Staub wrote:
Several OSDs of one of our clusters are down currently because RAM
usage has increased during the last days. Now it is more than we can
handle on some systems. Frequently OSDs get killed by the OOM
killer. Looking at "ceph daemon osd.$OSD_ID dump_mempools", it shows
that nearly all (about 8.5 GB) is taken by osd_pglog, e.g.
"osd_pglog": {
"items": 461859,
"bytes": 8445595868
},
We tried to reduce it, with "osd memory target" and even with
"bluestore cache autotune = false" (together with "bluestore cache
size hdd"), but there was no effect at all.
I remember the pglog_hardlimit parameter, but that is already set by
default with Nautilus I read. I.e. this is on Nautilus, 14.2.8.
Is there a way to limit this pglog memory?
Cheers
Harry
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
