Thanks to both of your suggestions, we had some progression on this issue.
I tuned vm.min_free_kbytes to 16GB and raised vm.vfs_cache_pressure to 200, and I did observe that the OS keep releasing cache while the OSDs want more and more memory.
OK. Now we are going to reproduce the hanging issue.
1. set the cluster with noup flag
2. restart all ceph-osd process (then we can see all OSDs are down from ceph monitor)
3. unset noup flag
As expected the OSDs started to consume memory, and eventually the kernel still hanged without response.
Therefore I learned to gather the vmcore and tried to investigate further as Brad advised.
The vmcore dump file was unbeliviably huge -- about 6 GB per dump. However it's helpful that we quickly found the following abnormal things:
1. The memory was exhausted as expected.
crash> kmem -i
PAGES TOTAL PERCENTAGE
TOTAL MEM 63322527 241.6 GB ----
FREE 676446 2.6 GB 1% of TOTAL MEM
USED 62646081 239 GB 98% of TOTAL MEM
SHARED 621336 2.4 GB 0% of TOTAL MEM
BUFFERS 47307 184.8 MB 0% of TOTAL MEM
CACHED 376205 1.4 GB 0% of TOTAL MEM
SLAB 455400 1.7 GB 0% of TOTAL MEM
TOTAL SWAP 4887039 18.6 GB ----
SWAP USED 3855938 14.7 GB 78% of TOTAL SWAP
SWAP FREE 1031101 3.9 GB 21% of TOTAL SWAP
COMMIT LIMIT 36548302 139.4 GB ----
COMMITTED 92434847 352.6 GB 252% of TOTAL LIMIT
2. Each OSD used a lot of memory. (We have only total 256 GB RAM but there are 90 OSDs in a node)
# Find 10 largest memory consumption processes
crash> ps -G | sed 's/>//g' | sort -k 8,8 -n | awk '$8 ~ /[0-9]/{ $8 = $8/1024" MB"; print}' | tail -10
100864 1 12 ffff883a43e1b700 IN 1.1 7484884 2973.33 MB ceph-osd
87400 1 27 ffff8838538ae040 IN 1.1 7557500 3036.92 MB ceph-osd
108126 1 22 ffff882bcca91b80 IN 1.2 7273068 3045.8 MB ceph-osd
39787 1 28 ffff883f468ab700 IN 1.2 7300756 3067.88 MB ceph-osd
44861 1 20 ffff883cf9250000 IN 1.2 7327496 3067.89 MB ceph-osd
30486 1 23 ffff883f59e1c4c0 IN 1.2 7332828 3083.58 MB ceph-osd
125239 1 15 ffff882687018000 IN 1.2 6965560 3103.36 MB ceph-osd
123807 1 19 ffff88275d90ee00 IN 1.2 7314484 3173.48 MB ceph-osd
116445 1 1 ffff882863926e00 IN 1.2 7279040 3269.09 MB ceph-osd
94442 1 0 ffff882ed2d01b80 IN 1.3 7566148 3418.69 MB ceph-osd
3. The excessive amount of message threads.
crash> ps | grep ms_pipe_read | wc -l
144112
crash> ps | grep ms_pipe_write | wc -l
146692
Totally up to 290k threads in ms_pipe_*.
4. Several tries we had, and we luckily got some memory profiles before oom killer started to work.
# Parse the smaps of a ceph-osd process by parse_smaps.py
root@ceph2:~# ./parse_smaps.py /proc/198557/smaps
===============================================================================
Private Private Shared Shared
Clean + Dirty + Clean + Dirty = Total : library
===============================================================================
16660 kB + 5804548 kB + 0 kB + 0 kB = 5821208 kB : [heap]
40 kB + 92 kB + 7640 kB + 0 kB = 7772 kB : ceph-osd
56 kB + 2472 kB + 0 kB + 0 kB = 2528 kB : [anonymous]
2084 kB + 0 kB + 0 kB + 0 kB = 2084 kB : 007656.ldb
2080 kB + 0 kB + 0 kB + 0 kB = 2080 kB : 007657.ldb
2080 kB + 0 kB + 0 kB + 0 kB = 2080 kB : 007653.ldb
2080 kB + 0 kB + 0 kB + 0 kB = 2080 kB : 007658.ldb
2080 kB + 0 kB + 0 kB + 0 kB = 2080 kB : 011125.ldb
2076 kB + 0 kB + 0 kB + 0 kB = 2076 kB : 009607.ldb
2072 kB + 0 kB + 0 kB + 0 kB = 2072 kB : 011127.ldb
2072 kB + 0 kB + 0 kB + 0 kB = 2072 kB : 011126.ldb
0 kB + 24 kB + 1636 kB + 0 kB = 1660 kB : libc-2.23.so
0 kB + 0 kB + 1060 kB + 0 kB = 1060 kB : libec_lrc.so
4 kB + 28 kB + 1024 kB + 0 kB = 1056 kB : libstdc++.so.6.0.21
996 kB + 0 kB + 0 kB + 0 kB = 996 kB : 011168.ldb
908 kB + 0 kB + 0 kB + 0 kB = 908 kB : 009608.ldb
840 kB + 0 kB + 0 kB + 0 kB = 840 kB : 007648.ldb
0 kB + 0 kB + 812 kB + 0 kB = 812 kB : libcls_rgw.so
0 kB + 0 kB + 716 kB + 0 kB = 716 kB : libcls_refcount.so
684 kB + 0 kB + 0 kB + 0 kB = 684 kB : 011128.ldb
0 kB + 0 kB + 552 kB + 0 kB = 552 kB : libm-2.23.so
0 kB + 0 kB + 472 kB + 0 kB = 472 kB : libec_jerasure_sse4.so
4 kB + 0 kB + 372 kB + 0 kB = 376 kB : libfreebl3.so
8 kB + 4 kB + 356 kB + 0 kB = 368 kB : libnss3.so
0 kB + 12 kB + 352 kB + 0 kB = 364 kB : libleveldb.so.1.18
0 kB + 0 kB + 296 kB + 0 kB = 296 kB : libec_jerasure.so
4 kB + 4 kB + 224 kB + 0 kB = 232 kB : libsoftokn3.so
8 kB + 0 kB + 208 kB + 0 kB = 216 kB : libnspr4.so
8 kB + 0 kB + 196 kB + 0 kB = 204 kB : libec_isa.so
0 kB + 8 kB + 196 kB + 0 kB = 204 kB : libtcmalloc.so.4.2.6
0 kB + 8 kB + 152 kB + 0 kB = 160 kB : ld-2.23.so
4 kB + 4 kB + 136 kB + 0 kB = 144 kB : libboost_thread.so.1.58.0
4 kB + 0 kB + 132 kB + 0 kB = 136 kB : ibnssutil3.so
......
===============================================================================
37736 kB + 5807224 kB + 18128 kB + 0 kB = 5863088 kB : Total
5. Heap profiler by Ceph.
root@ceph2:~# ceph tell osd.163 heap stats
osd.163 tcmalloc heap stats:------------------------------------------------
MALLOC: 5861094560 ( 5589.6 MiB) Bytes in use by application
MALLOC: + 0 ( 0.0 MiB) Bytes in page heap freelist
MALLOC: + 38945176 ( 37.1 MiB) Bytes in central cache freelist
MALLOC: + 13279168 ( 12.7 MiB) Bytes in transfer cache freelist
MALLOC: + 96438792 ( 92.0 MiB) Bytes in thread cache freelists
MALLOC: + 25817248 ( 24.6 MiB) Bytes in malloc metadata
MALLOC: ------------
MALLOC: = 6035574944 ( 5756.0 MiB) Actual memory used (physical + swap)
MALLOC: + 35741696 ( 34.1 MiB) Bytes released to OS (aka unmapped)
MALLOC: ------------
MALLOC: = 6071316640 ( 5790.1 MiB) Virtual address space used
MALLOC:
MALLOC: 357627 Spans in use
MALLOC: 89 Thread heaps in use
MALLOC: 8192 Tcmalloc page size
------------------------------------------------
6. google-pprof the heap dump
Total: 1916.6 MB
1036.9 54.1% 54.1% 1036.9 54.1% ceph::buffer::create_aligned
313.9 16.4% 70.5% 313.9 16.4% ceph::buffer::list::append@a78c00
220.0 11.5% 82.0% 220.0 11.5% std::_Rb_tree::_M_emplace_hint_unique
130.0 6.8% 88.7% 130.0 6.8% leveldb::ReadBlock
129.8 6.8% 95.5% 129.8 6.8% std::vector::_M_default_append
22.1 1.2% 96.7% 53.4 2.8% PG::add_log_entry
7.4 0.4% 97.1% 7.4 0.4% ceph::buffer::list::crc32c
7.0 0.4% 97.4% 7.0 0.4% ceph::log::Log::create_entry
5.1 0.3% 97.7% 5.1 0.3% OSD::get_tracked_conf_keys
4.7 0.2% 97.9% 4.8 0.2% Pipe::Pipe
3.3 0.2% 98.1% 9.0 0.5% decode_message
3.2 0.2% 98.3% 6.7 0.3% SimpleMessenger::add_accept_pipe
3.2 0.2% 98.4% 4.5 0.2% OSD::_make_pg
...
We have some hypotheses after discussion:
1. We observed that the number of connection (counted by `netstat -ant | grep ESTABLISHED | wc -l`) rises rapidly along with the average memory used by ceph-osd, especially [heap] section.
> Is there any relation between memory usage and the number of network connection?
2. After unset noup flag, the number of connection bursts to over 200k in few seconds.
> We have an EC pool created with k=17, m=3. Is the large combination of (k,m) responsible for these connections?
> We have average 300 pgs per OSD in the crash experiment. Is the high pgs per OSD responsible for these connections?
3. With simple messenger we forked two message threads for single network connection.
> We think 290k message threads in the same time are hard to work normally and efficiently.
> Will it be better with async messenger? Tried with async messenger, we saw threads decreased but the number of network connection still high and the kernel hang issue continued.
Now we are still struggling with this problem.
Please kindly instruct us if you have any directions.
Craig Chi
Hi,
I didn’t so the maths, so maybe 7GB isn’t worth tuning for, although every little helps ;-)
I don’t believe peering or recovery should effect this value, but other things will consume memory during recovery, but I’m not aware if this can be limited or tuned.
Yes, the write and read cache’s will consume memory and may limit Linux’s ability to react quickly enough in tight memory conditions. I believe you can be in a state where it looks like you have more memory potentially available than actually is usable at that point in time. The min_free_bytes can help here.
From: Craig Chi [mailto:craigchi@xxxxxxxxxxxx]
Sent: 25 November 2016 01:46
To: Brad Hubbard <bhubbard@xxxxxxxxxx>
Cc: Nick Fisk <nick@xxxxxxxxxx>; Ceph Users <ceph-users@xxxxxxxxxxxxxx>
Subject: Re: Ceph OSDs cause kernel unresponsive
Hi Nick,
I have seen the report before, if I understand correctly, the osd_map_cache_size generally introduces a fixed amount of memory usage. We are using the default value of 200, and a single osd map I got from our cluster is 404KB.
That is totally 404KB * 200 * 90 (osds) = about 7GB on each node.
Will the memory consumption generated by this factor become larger when unstably peering or recovering? If not, we still need to find the root cause of why free memory drops without control.
Does anyone know that what is the relation between filestore or journal configurations and the OSD's memory consumption? Is it possible that the filestore queue or journal queue occupy huge memory pages and cause filesystem cache hard to release (and result in oom)?
At last, about nobarrier, I fully knew the consequence and is seriously testing on this option. Sincerely appreciate your kindness and useful suggestions.
Sincerely,
Craig ChiOn 2016-11-25 07:23, Brad Hubbard <bhubbard@xxxxxxxxxx> wrote:
Two of these appear to be hung task timeouts and the other is an invalid opcode.
There is no evidence here of memory exhaustion (although it remains to be seen whether this is a factor but I'd expect to see evidence of shrinker activity in the stacks) and I would speculate the increased memory utilisation is due to the issues with the OSD tasks.
I would suggest that the next step here is to work out specifically why the invalid opcode happened and/or why kernel tasks are hanging for > 120 seconds.
To do that you may need to capture a vmcore and analyse it and/or engage your kernel support team to investigate further.
On Fri, Nov 25, 2016 at 8:26 AM, Nick Fisk <nick@xxxxxxxxxx> wrote:
There’s a couple of things you can do to reduce memory usage by limiting the number of OSD maps each OSD stores, but you will still be pushing up against the limits of the ram you have available. There is a Cern 30PB test (should be on google) which gives some details on some of the settings, but quite a few are no longer relevant in jewel.
Once other thing, I saw you have nobarrier set on mount options. Please please please understand the consequences of this option!!!!
From: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] On Behalf Of Craig Chi
Sent: 24 November 2016 10:37
To: Nick Fisk <nick@xxxxxxxxxx>
Cc: ceph-users@xxxxxxxxxxxxxx
Subject: Re: Ceph OSDs cause kernel unresponsive
Hi Nick,
Thank you for your helpful information.
I knew that Ceph recommends 1GB/1TB RAM, but we are not going to change the hardware architecture now.
Are there any methods to set the resource limit one OSD can consume?
And for your question, we currently set system configuration as:
vm.swappiness=10
kernel.pid_max=4194303
fs.file-max=26234859
vm.zone_reclaim_mode=0
vm.vfs_cache_pressure=50
vm.min_free_kbytes=4194303
I would try to configure vm.min_free_kbytes larger and test.
I will be grateful if anyone has the experience of how to tune these values for Ceph.
Sincerely,
Craig Chi
On 2016-11-24 17:48, Nick Fisk <nick@xxxxxxxxxx> wrote:
Hi Craig,
From: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] On Behalf Of Craig Chi
Sent: 24 November 2016 08:34
To: ceph-users@xxxxxxxxxxxxxx
Subject: Ceph OSDs cause kernel unresponsive
Hi Cephers,
We have encountered kernel hanging issue on our Ceph cluster. Just like http://imgur.com/a/U2Flz , http://imgur.com/a/lyEko or http://imgur.com/a/IGXdu .
We believed it is caused by out of memory, because we observed that when OSDs went crazy, the available memory of each node were decreasing rapidly (from 50% available to lower than 10%). Then the node running Ceph OSD became unresponsive with console showing hung_task_timout or slab_out_of_memory, etc. The only thing we can do then is hard reset the unit.
It is hard to predict when the kernel hanging issue will happen. In my past experiences, it usually happened after a long term benchmark procedure, and followed by a manual trigger like 1) reboot a node 2) restart all OSDs 3) modify CRUSH map.
Currently the cluster is back to normal, but we want to figure out the root cause to avoid happening again. We think the high values of ceph.conf are pretty suspicous, but without code tracing we are hard to realize the impact of the values and the memory consumption.
Many thanks if you have any suggestions.
I think you are probably running out of memory, 90x8TB disks is 720Tb of storage, that will need a lot of ram to run and also the fact that the problems occur when PG’s start moving around after a node failure also suggests this.
Have you adjusted your vm.vfs_cache_pressure?
You might also want to try setting vm.min_free_kbytes to 8-16GB to try and keep some memory free and avoid fragmentation.
=================================================================================
Following is our ceph cluster architecture:
OS: Ubuntu 16.04.1 LTS (4.4.0-31-generic #50-Ubuntu x86_64 GNU/Linux)
Ceph: Jewel 10.2.3
3 Ceph Monitors running on 3 dedicated machines
630 Ceph OSDs running on 7 storage machines (each machine has 256GB RAM and 90 units of 8TB hard drives)
There are 4 pools with following settings:
vms 512 pg x 3 replica
images 512 pg x 3 replica
volumes 8192 pg x 3 replica
objects 4096 pg x (17,3) erasure code profile
==> average 173.92 pgs per OSD
We tuned our ceph.conf by referencing many performance tuning resources online ( mainly from slide 38 of https://goo.gl/Idkh41 )
[global]
osd pool default pg num = 4096
osd pool default pgp num = 4096
err to syslog = true
log to syslog = true
osd pool default size = 3
max open files = 131072
fsid = 1c33bf75-e080-4a70-9fd8-860ff216f595
osd crush chooseleaf type = 1
[mon.mon1]
host = mon1
mon addr = 172.20.1.2
[mon.mon2]
host = mon2
mon addr = 172.20.1.3
[mon.mon3]
host = mon3
mon addr = 172.20.1.4
[mon]
mon osd full ratio = 0.85
mon osd nearfull ratio = 0.7
mon osd down out interval = 600
mon osd down out subtree limit = host
mon allow pool delete = true
mon compact on start = true
[osd]
public_network = 172.20.3.1/21
cluster_network = 172.24.0.1/24
osd disk threads = 4
osd mount options xfs = rw,noexec,nodev,noatime,nodiratime,nobarrier,inode64,logbsize=256k
osd crush update on start = false
osd op threads = 20
osd mkfs options xfs = -f -i size=2048
osd max write size = 512
osd mkfs type = xfs
osd journal size = 5120
filestore max inline xattrs = 6
filestore queue committing max bytes = 1048576000
filestore queue committing max ops = 5000
filestore queue max bytes = 1048576000
filestore op threads = 32
filestore max inline xattr size = 254
filestore max sync interval = 15
filestore min sync interval = 10
journal max write bytes = 1048576000
journal max write entries = 1000
journal queue max ops = 3000
journal queue max bytes = 1048576000
ms dispatch throttle bytes = 1048576000
Sincerely,
Craig Chi
Sent from Synology MailPlus
Sent from Synology MailPlus
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
--Cheers,
Brad
Sent from Synology MailPlus
Sent from Synology MailPlus
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
