[Linux-cluster] GFS hangs, nodes die
Sebastian Walter
sebastian.walter at fu-berlin.de
Mon Aug 20 08:24:20 UTC 2007
Hi Marc,
thanks for your answer. Yesterday night, I set the 'gfs_tool settune
glock_purge 50'. After putting some stress on the cluster, one node now
failed, while gfs is still running, but very, very slow (maybe it will
collapse soon). /proc/cluster/services tells me that the nodes are all
still ok, and the failed doesn't get fenced (normally, fencing works
well in the environment). But I can't ssh into the node, and on the
console I can't login (even if it still let's me type it doesn't react).
The main jobs of the cluster went to 'D' state.
This is what `gfs_tool counters /global/home` gives me (this command
also hang, but after restarting the failed node it worked):
[root at dtm ~]# gfs_tool counters /global/home
locks 23532
locks held 10774
freeze count 0
incore inodes 10764
metadata buffers 4
unlinked inodes 1
quota IDs 2
incore log buffers 0
log space used 0.10%
meta header cache entries 0
glock dependencies 0
glocks on reclaim list 0
log wraps 2
outstanding LM calls 0
outstanding BIO calls 0
fh2dentry misses 0
glocks reclaimed 421095
glock nq calls 2196642
glock dq calls 2051456
glock prefetch calls 171529
lm_lock calls 345420
lm_unlock calls 155518
lm callbacks 515665
address operations 4431
dentry operations 377484
export operations 0
file operations 535012
inode operations 892178
super operations 1196275
vm operations 0
block I/O reads 126680
block I/O writes 133808
Any suggestions?
Regards,
Sebastian
Marc Grimme wrote:
> Hi Sebastian,
> you might also want to have a look at here:
> http://www.open-sharedroot.org/Members/marc/blog/blog-on-gfs/
> I collected some information about the problem you've hit (It must be that
> problem).
> Next time you should also look at the console of every node. You should see
> some intesting messages before there.
> Use the glock_purge gfs_tool option it will help and always keep a look on the
> gfs_tool counters and there on the locks.
>
> BTW: the unable to obtain lock is only the rgmanager complaining about not
> being able to obtain a lock and as side effect. The problem is that a new
> lockid cannot be got within time.
>
> Regards Marc.
> On Sunday 19 August 2007 11:53:39 you wrote:
>
>> Hi Marc!
>>
>> Thanks for your help. As I restarted everything now, I can't check this.
>> I will do when it's crahsing again (I will do some tests now). I
>> realised that one node did hang with kernel panic. Attached is the
>> screenshot.
>>
>> regards
>> sebastian
>>
>> Marc Grimme wrote:
>>
>>> Hello Sebastian,
>>> what do gfs_tool counters on the fs tell you?
>>> And ps axf? Do you have a lot of "D" processes?
>>> Regards Marc.
>>>
>>> On Sunday 19 August 2007 02:06:30 Sebastian Walter wrote:
>>>
>>>> Dear list,
>>>>
>>>> this is the tragical story of my cluster running rhel/csgfs 4u5: the
>>>> cluster in generally is running fine, but when I increase the load to a
>>>> certain level (heavy I/O), it collapses. About 20% of the nodes do crash
>>>> (not reacting any more, but no sign of kernel panic), the others can't
>>>> access the gfs resource.
>>>> Gfs is set up as a rgmanager service with failover domain for each node
>>>> (same problem also exists when mounting via /etc/fstab).
>>>>
>>>> Who is willing to provide a happy end?
>>>>
>>>> Thanks, Sebastian
>>>> **
>>>>
>>>> This is what /var/log/messages gives me (on nearly all nodes):
>>>> Aug 18 04:39:06 compute-0-2 clurgmgrd[4225]: <err> #49: Failed getting
>>>> status for RG gfs-2
>>>> and e.g.
>>>> Aug 18 04:45:38 compute-0-6 clurgmgrd[9074]: <err> #50: Unable to obtain
>>>> cluster lock: Connection timed out
>>>>
>>>> [root at compute-0-3 ~]# cat /proc/cluster/status
>>>> Protocol version: 5.0.1
>>>> Config version: 53
>>>> Cluster name: dtm
>>>> Cluster ID: 741
>>>> Cluster Member: Yes
>>>> Membership state: Cluster-Member
>>>> Nodes: 10
>>>> Expected_votes: 11
>>>> Total_votes: 10
>>>> Quorum: 6
>>>> Active subsystems: 8
>>>> Node name: compute-0-3
>>>> Node ID: 4
>>>> Node addresses: 10.1.255.252
>>>>
>>>> [root at compute-0-6 ~]# cat /proc/cluster/services
>>>> Service Name GID LID State
>>>> Code Fence Domain: "default" 3 2 recover
>>>> 4 - [1 2 6 10 9 8 3 7 4 11]
>>>> DLM Lock Space: "clvmd" 7 3 recover 0 -
>>>> [1 2 6 10 9 8 3 7 4 11]
>>>> DLM Lock Space: "Magma" 12 5 recover 0 -
>>>> [1 2 6 10 9 8 3 7 4 11]
>>>> DLM Lock Space: "homeneu" 17 6 recover 0 -
>>>> [10 9 8 7 2 3 6 4 1 11]
>>>> GFS Mount Group: "homeneu" 18 7 recover 0 -
>>>> [10 9 8 7 2 3 6 4 1 11]
>>>> User: "usrm::manager" 11 4 recover 0 -
>>>> [1 2 6 10 9 8 3 7 4 11]
>>>>
>>>> [root at compute-0-10 ~]# cat /proc/cluster/dlm_stats
>>>> DLM stats (HZ=1000)
>>>>
>>>> Lock operations: 4036
>>>> Unlock operations: 2001
>>>> Convert operations: 1862
>>>> Completion ASTs: 7898
>>>> Blocking ASTs: 52
>>>>
>>>> Lockqueue num waittime ave
>>>> WAIT_RSB 3778 28862 7
>>>> WAIT_CONV 75 482 6
>>>> WAIT_GRANT 2171 7235 3
>>>> WAIT_UNLOCK 153 1606 10
>>>> Total 6177 38185 6
>>>>
>>>> [root at compute-0-10 ~]# cat /proc/cluster/sm_debug
>>>> sevent state 7
>>>> 02000012 sevent state 9
>>>> 00000003 remove node 5 count 10
>>>> 01000011 remove node 5 count 10
>>>> 0100000c remove node 5 count 10
>>>> 01000007 remove node 5 count 10
>>>> 02000012 remove node 5 count 10
>>>> 0300000b remove node 5 count 10
>>>> 00000003 recover state 0
>>>>
>>>>
>>>>
>>>> --
>>>> Linux-cluster mailing list
>>>> Linux-cluster at redhat.com
>>>> https://www.redhat.com/mailman/listinfo/linux-cluster
>>>>
>
>
>
>
More information about the Linux-cluster
mailing list