Hi,
We are back to our MDS problem.
This time we used the kernel client to connect to the cluster. We
started again filling the cluster with rsync, and got problems after
about 48h. The MDS crashed again with core dump like before.
I removed the client connection, and restarted the MDS's. The first
MDS got killed by oom-killer when it used 8G of RAM. The backup MDS
took over, and I stopped the OSDs(with noout set) on that host so
there was more RAM for testing the MDS. The MDS started again to fill
the memory, this time to 30G. (We have 32G)
Meanwhile, I saw some interesting rados statistics:
The metadata pool was heavily decreasing in KB, and increasing in
object count.
pool name category KB objects clones
degraded unfound rd rd KB wr wr KB
data - 17707360279 11641299 0
11667979 0 70364 0 15904215
17711309640
metadata - 95833067 405395 0
484371 0 273320 925385352 12283908
271184622
Tue Apr 15 12:55:34 2014
pool name category KB objects clones
degraded unfound rd rd KB wr wr KB
data - 17707360279 11667145 0
11719671 0 95576 0 17619242
17711309640
metadata - 49677494 494630 0
662841 0 294946 1013866123 12748893
271914066
After oom-killer killed ceph-mds again:
Out of memory: Kill process 40776 (ceph-mds) score 808 or sacrifice
child Killed process 40776, UID 0, (ceph-mds) total-vm:31382436kB,
anon-rss:29960904kB, file-rss:20kB python[42731] trap invalid opcode
ip:7f62ddd8ffc1 sp:7f62ccef6488 error:0
this reduction was still going on (no clients connected):
Tue Apr 15 13:20:24 2014
pool name category KB objects clones
degraded unfound rd rd KB wr wr KB
data - 17707360279 11686446 0
11758273 0 95576 0 18447710
17711309640
metadata - 26526902 523234 0
720049 0 294946 1013866123 12886549
271914355
Tue Apr 15 13:34:49 2014
pool name category KB objects clones
degraded unfound rd rd KB wr wr KB
data - 17707360279 11707425 0
11800231 0 95576 0 18989100
17711309640
metadata - 12612790 543214 0
760009 0 294946 1013866123 12977285
271914524
The MDS didn't crash anymore and stopped increasing RAM at 20G
Until it eventually stabilized:
Tue Apr 15 13:47:35 2014
pool name category KB objects clones
degraded unfound rd rd KB wr wr KB
data - 17707360279 11720750 0
11826881 0 95576 0 19346286
17711309640
metadata - 126985 579061 0
831703 0 294946 1013866123 13146061
272057126
The MDS is now only using <200m and idle
Does this looks familiar to someone?
It looks like the MDS is unable to process the requests, and therefore
tried to allocate more memory until it resolved?
Thanks!
Kenneth
----- Message from Stijn De Weirdt <[email protected]> ---------
Date: Fri, 04 Apr 2014 20:31:34 +0200
From: Stijn De Weirdt <[email protected]>
Subject: Re: [ceph-users] ceph 0.78 mon and mds crashing (bus error)
To: "Yan, Zheng" <[email protected]>
Cc: [email protected]
hi yan,
(taking the list in CC)
On 04/04/2014 04:44 PM, Yan, Zheng wrote:
On Thu, Apr 3, 2014 at 2:52 PM, Stijn De Weirdt
<[email protected]> wrote:
hi,
latest pprof output attached.
this is no kernel client, this is ceph-fuse on EL6. starting the
mds without
any ceph-fuse mounts works without issue. mounting ceph-fuse
afterwards also
works fine. simple filesystem operations work as expected.
we'll check the state of the fuse mount via fusectl once we've reproduced
the issue.
One dark side of fuse client is that there is no way to purge inodes
in kernel inode caches. Each cached inode in kernel pins corresponding
inode in ceph-mds. I think that's why ceph-fuse and ceph-mds use so
many memory. Could you try executing "echo 3 >
/proc/sys/vm/drop_caches" periodically on the ceph-fuse mount machine.
thanks a lot for the suggestion. that night very well explain the
issue. we're currently trying looking into other issues that we
tought was related (extremely low mdtest performance), but we will
retry your suggestion when we start to reproduce the condition
stijn
Regards
Yan, Zheng
stijn
On 04/03/2014 02:39 AM, Yan, Zheng wrote:
which version of kernel client did you use? please send out context of
client node's /sys/kernel/debug/ceph/*/caps when the MDS uses lots
memory.
Regards
Yan, Zheng
On Thu, Apr 3, 2014 at 2:58 AM, Stijn De Weirdt <[email protected]>
wrote:
wow. kudos for integrating this in ceph. more projects should do it like
that!
anyway, in attachement a gzipped ps file. heap is at 4.4GB, top reports
6.5GB mem usage.
care to point out what to look for? i'll send a new one when the usage is
starting to cause swapping.
thanks,
stijn
On 04/02/2014 06:35 PM, Gregory Farnum wrote:
Did you see
http://ceph.com/docs/master/rados/troubleshooting/memory-profiling?
That should have what you need to get started, although you'll also
need to learn the basics of using the heap analysis tools elsewhere.
-Greg
Software Engineer #42 @ http://inktank.com | http://ceph.com
On Wed, Apr 2, 2014 at 9:23 AM, Stijn De Weirdt
<[email protected]>
wrote:
hi,
1) How big and what shape the filesystem is. Do you have some
extremely large directory that the MDS keeps trying to load and then
dump?
anyway to extract this from the mds without having to start it? as it
was an rsync operation, i can try to locate possible candidates on the
source filesystem, but what would be considered "large"?
total number of files 13M, spread over 800k directories, but it's
unclear
how far the sync was at time of failing. i've not found a good way to
for
directories with lots of files and/or subdirs.
2) Use tcmalloc's heap analyzer to see where all the memory is being
allocated.
we'll giv ethat a try
i run ceph-mds with HEAPCHECK=normal (via the init script), but how can
we
"stop" mds without killing it? the heapchecker only seems to dump at
the
end
of a run, maybe there's a way to have intermediate dump like valgrind,
but
the documentation is not very helpful.
stijn
3) Look through the logs for when the beacon fails (the first of
"mds.0.16 is_laggy 600.641332 > 15 since last acked beacon") and see
if there's anything tell-tale going on at the time.
anything in particular we should be looking for?
the log goes as follows:
mds starts around 11:43
...
2014-04-01 11:44:23.658583 7ffec89c6700 1 mds.0.server
reconnect_clients -- 1 sessions
2014-04-01 11:44:41.212488 7ffec89c6700 0 log [DBG] : reconnect by
client.4585 10.141.8.199:0/3551 after 17.553854
2014-04-01 11:44:45.692237 7ffec89c6700 1 mds.0.10 reconnect_done
2014-04-01 11:44:45.996384 7ffec89c6700 1 mds.0.10 handle_mds_map i
am now mds.0.10
2014-04-01 11:44:45.996388 7ffec89c6700 1 mds.0.10 handle_mds_map
state change up:reconnect --> up:rejoin
2014-04-01 11:44:45.996390 7ffec89c6700 1 mds.0.10 rejoin_start
2014-04-01 11:49:53.158471 7ffec89c6700 1 mds.0.10
rejoin_joint_start
then lots (4667 lines) of
2014-04-01 11:50:10.237035 7ffebc844700 0 -- 10.141.8.180:6837/55117
10.141.8.180:6789/0 pipe(0x38a7da00 sd=104 :41115 s=4 pgs=0 cs=0
l=1 c=0x6513e8840).connect got RESETSESSION but no longer connecting
with one intermediate
2014-04-01 11:51:50.181354 7ffebcf4b700 0 -- 10.141.8.180:6837/55117
10.141.8.180:6789/0 pipe(0x10e282580 sd=103 :0 s=1 pgs=0 cs=0 l=1
c=0xc77d5ee0).fault
then sudden change
2014-04-01 11:57:30.440554 7ffebcd49700 0 -- 10.141.8.180:6837/55117
10.141.8.182:6789/0 pipe(0xa1534100 sd=104 :48176 s=4 pgs=0 cs=0
l=1 c=0xd99b11e0).connect got RESETSESSION but no longer connecting
2014-04-01 11:57:30.722607 7ffebec68700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0x1ce98a00 sd=104 :48235 s=1 pgs=0 cs=0
l=1 c=0xc48a3f40).connect got BADAUTHORIZER
2014-04-01 11:57:30.722669 7ffebec68700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0x1ce98a00 sd=104 :48235 s=1 pgs=0 cs=0
l=1 c=0xc48a3f40).connect got BADAUTHORIZER
2014-04-01 11:57:30.722885 7ffebec68700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0x1ce98a00 sd=57 :48237 s=1 pgs=0 cs=0 l=1
c=0xc48a3f40).connect got BADAUTHORIZER
2014-04-01 11:57:30.722945 7ffebec68700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0x1ce98a00 sd=57 :48237 s=1 pgs=0 cs=0 l=1
c=0xc48a3f40).connect got BADAUTHORIZER
followed by lots of
2014-04-01 11:57:30.738562 7ffebec68700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0xead9fa80 sd=57 :0 s=1 pgs=0 cs=0 l=1
c=0x10e5d280).fault
with sporadic
2014-04-01 11:57:32.431219 7ffebeb67700 0 -- 10.141.8.180:6837/55117
10.141.8.182:6789/0 pipe(0xef85cd80 sd=103 :41218 s=4 pgs=0 cs=0
l=1 c=0x130590dc0).connect got RESETSESSION but no longer connecting
until the dmup
2014-04-01 11:59:27.612850 7ffebea66700 0 -- 10.141.8.180:6837/55117
10.141.8.181:6789/0 pipe(0xe3036400 sd=103 :0 s=1 pgs=0 cs=0 l=1
c=0xa7be300).fault
2014-04-01 11:59:27.639009 7ffec89c6700 -1 mds/CDir.cc: In function
'void CDir::_omap_fetched(ceph::bufferlist&,
std::map<std::basic_string<char, std::char_traits<char>,
std::allocator\
<char> >, ceph::buffer::list, std::less<std::basic_string<char,
std::char_traits<char>, std::allocator<char> > >,
std::allocator<std::pair<const std::basic_string<char,
std::char_trait\
s<char>, std::allocator<char> >, ceph::buffer::list> > >&, const
std::string&, int)' thread 7ffec89c6700 time 2014-04-01
11:59:27.620684
mds/CDir.cc: 1474: FAILED assert(r == 0 || r == -2 || r == -61)
ceph version 0.78 (f6c746c314d7b87b8419b6e584c94bfe4511dbd4)
1: (CDir::_omap_fetched(ceph::buffer::list&, std::map<std::string,
ceph::buffer::list, std::less<std::string>,
std::allocator<std::pair<std::string const, ceph::buffer::list> > >&,
st\
d::string const&, int)+0x4d71) [0x77c3c1]
2: (Context::complete(int)+0x9) [0x56bb79]
3: (Objecter::handle_osd_op_reply(MOSDOpReply*)+0x11a6) [0x806dd6]
4: (MDS::handle_core_message(Message*)+0x9c7) [0x5901d7]
5: (MDS::_dispatch(Message*)+0x2f) [0x59028f]
6: (MDS::ms_dispatch(Message*)+0x1ab) [0x591d4b]
7: (DispatchQueue::entry()+0x582) [0x902072]
8: (DispatchQueue::DispatchThread::entry()+0xd) [0x85ef4d]
9: /lib64/libpthread.so.0() [0x34c36079d1]
10: (clone()+0x6d) [0x34c32e8b6d]
NOTE: a copy of the executable, or `objdump -rdS <executable>` is
needed to interpret this.
--- begin dump of recent events ---
-10000> 2014-04-01 11:59:27.137779 7ffec89c6700 5 mds.0.10
initiating
monitor reconnect; maybe we're not the slow one
-9999> 2014-04-01 11:59:27.137787 7ffec89c6700 10
monclient(hunting):
_reopen_session rank -1 name
-9998> 2014-04-01 11:59:27.137790 7ffec89c6700 1 --
10.141.8.180:6837/55117 mark_down 0x1326be460 -- 0x2cae5f00
-9997> 2014-04-01 11:59:27.137807 7ffebf672700 2 --
10.141.8.180:6837/55117 >> 10.141.8.181:6789/0 pipe(0x2cae5f00 sd=57
:0 s=4 pgs=0 cs=0 l=1 c=0x1326be460).connect error 10.141.8.1\
81:6789/0, 104: Connection reset by peer
-9996> 2014-04-01 11:59:27.137819 7ffebf672700 3 --
10.141.8.180:6837/55117 >> 10.141.8.181:6789/0 pipe(0x2cae5f00 sd=57
:0 s=4 pgs=0 cs=0 l=1 c=0x1326be460).connect fault, but state\
= closed != connecting, stopping
-9995> 2014-04-01 11:59:27.137824 7ffec89c6700 10
monclient(hunting):
picked mon.ceph003 con 0x3af54ba0 addr 10.141.8.182:6789/0
-9994> 2014-04-01 11:59:27.137832 7ffec89c6700 10
monclient(hunting):
_send_mon_message to mon.ceph003 at 10.141.8.182:6789/0
-Greg
Software Engineer #42 @ http://inktank.com | http://ceph.com
On Wed, Apr 2, 2014 at 3:39 AM, Kenneth Waegeman
<[email protected]> wrote:
----- Message from Gregory Farnum <[email protected]> ---------
Date: Tue, 1 Apr 2014 09:03:17 -0700
From: Gregory Farnum <[email protected]>
Subject: Re: [ceph-users] ceph 0.78 mon and mds crashing (bus error)
To: "Yan, Zheng" <[email protected]>
Cc: Kenneth Waegeman <[email protected]>, ceph-users
<[email protected]>
On Tue, Apr 1, 2014 at 7:12 AM, Yan, Zheng <[email protected]>
wrote:
On Tue, Apr 1, 2014 at 10:02 PM, Kenneth Waegeman
<[email protected]> wrote:
After some more searching, I've found that the source of the
problem is
with
the mds and not the mon.. The mds crashes, generates a core dump
that
eats
the local space, and in turn the monitor (because of leveldb)
crashes.
The error in the mds log of one host:
2014-04-01 15:46:34.414615 7f870e319700 0 --
10.141.8.180:6836/13152 >>
10.141.8.180:6789/0 pipe(0x517371180 sd=54 :42439 s=4 pgs=0 cs=0
l=1
c=0x147ac780).connect got RESETSESSION but no longer connecting
2014-04-01 15:46:34.438792 7f871194f700 0 --
10.141.8.180:6836/13152 >>
10.141.8.180:6789/0 pipe(0x1b099f580 sd=8 :43150 s=4 pgs=0 cs=0
l=1
c=0x1fd44360).connect got RESETSESSION but no longer connecting
2014-04-01 15:46:34.439028 7f870e319700 0 --
10.141.8.180:6836/13152 >>
10.141.8.182:6789/0 pipe(0x13aa64880 sd=54 :37085 s=4 pgs=0 cs=0
l=1
c=0x1fd43de0).connect got RESETSESSION but no longer connecting
2014-04-01 15:46:34.468257 7f871b7ae700 -1 mds/CDir.cc: In
function
'void
CDir::_omap_fetched(ceph::bufferlist&,
std::map<std::basic_string<char,
std::char_traits<char>, std::allocator<char> >,
ceph::buffer::list,
std::less<std::basic_string<char, std::char_traits<char>,
std::allocator<char> > >, std::allocator<std::pair<const
std::basic_string<char, std::char_traits<char>,
std::allocator<char> >,
ceph::buffer::list> > >&, const std::string&, int)' thread
7f871b7ae700
time
2014-04-01 15:46:34.448320
mds/CDir.cc: 1474: FAILED assert(r == 0 || r == -2 || r == -61)
could you use gdb to check what is value of variable 'r' .
If you look at the crash dump log you can see the return value in
the
osd_op_reply message:
-1> 2014-04-01 15:46:34.440860 7f871b7ae700 1 --
10.141.8.180:6836/13152 <== osd.3 10.141.8.180:6827/4366 33077 ====
osd_op_reply(4179177 100001f2ef1.00000000 [omap-get-header
0~0,omap-get-vals 0~16] v0'0 uv0 ack = -108 (Cannot send after
transport endpoint shutdown)) v6 ==== 229+0+0 (958358678 0 0)
0x2cff7aa80 con 0x37ea3c0
-108, which is ESHUTDOWN, but we also use it (via the 108 constant,
I
think because ESHUTDOWN varies across platforms) as EBLACKLISTED.
So it looks like this is itself actually a symptom of another
problem
that is causing the MDS to get timed out on the monitor. If a core
dump is "eating the local space", maybe the MDS is stuck in an
infinite allocation loop of some kind? How big are your disks,
Kenneth? Do you have any information on how much CPU/memory the MDS
was using before this?
I monitored the mds process after restart:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
19215 root 20 0 6070m 5.7g 5236 S 778.6 18.1 1:27.54
ceph-mds
19215 root 20 0 7926m 7.5g 5236 S 179.2 23.8 2:44.39
ceph-mds
19215 root 20 0 12.4g 12g 5236 S 157.2 38.8 3:43.47
ceph-mds
19215 root 20 0 16.6g 16g 5236 S 144.4 52.0 4:15.01
ceph-mds
19215 root 20 0 19.9g 19g 5236 S 137.2 62.5 4:35.83
ceph-mds
19215 root 20 0 24.5g 24g 5224 S 136.5 77.0 5:04.66
ceph-mds
19215 root 20 0 25.8g 25g 2944 S 33.7 81.2 5:13.74
ceph-mds
19215 root 20 0 26.0g 25g 2916 S 24.6 81.7 5:19.07
ceph-mds
19215 root 20 0 26.1g 25g 2916 S 13.0 82.1 5:22.16
ceph-mds
19215 root 20 0 27.7g 26g 1856 S 100.0 85.8 5:36.46
ceph-mds
Then it crashes. I changed the core dump location out of the root
fs,
the
core dump is indeed about 26G
My disks:
Filesystem Size Used Avail Use% Mounted on
/dev/sda2 9.9G 2.9G 6.5G 31% /
tmpfs 16G 0 16G 0% /dev/shm
/dev/sda1 248M 53M 183M 23% /boot
/dev/sda4 172G 61G 112G 35% /var/lib/ceph/log/sda4
/dev/sdb 187G 61G 127G 33% /var/lib/ceph/log/sdb
/dev/sdc 3.7T 1.7T 2.0T 47% /var/lib/ceph/osd/sdc
/dev/sdd 3.7T 1.5T 2.2T 41% /var/lib/ceph/osd/sdd
/dev/sde 3.7T 1.4T 2.4T 37% /var/lib/ceph/osd/sde
/dev/sdf 3.7T 1.5T 2.3T 39% /var/lib/ceph/osd/sdf
/dev/sdg 3.7T 2.1T 1.7T 56% /var/lib/ceph/osd/sdg
/dev/sdh 3.7T 1.7T 2.0T 47% /var/lib/ceph/osd/sdh
/dev/sdi 3.7T 1.7T 2.0T 47% /var/lib/ceph/osd/sdi
/dev/sdj 3.7T 1.7T 2.0T 47% /var/lib/ceph/osd/sdj
/dev/sdk 3.7T 2.1T 1.6T 58% /var/lib/ceph/osd/sdk
/dev/sdl 3.7T 1.7T 2.0T 46% /var/lib/ceph/osd/sdl
/dev/sdm 3.7T 1.5T 2.2T 41% /var/lib/ceph/osd/sdm
/dev/sdn 3.7T 1.4T 2.3T 38% /var/lib/ceph/osd/sdn
-Greg
Software Engineer #42 @ http://inktank.com | http://ceph.com
----- End message from Gregory Farnum <[email protected]> -----
--
Met vriendelijke groeten,
Kenneth Waegeman
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
----- End message from Stijn De Weirdt <[email protected]> -----
--
Met vriendelijke groeten,
Kenneth Waegeman
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
