On Fri, Jul 1, 2016 at 6:59 PM, John Spray <jspray@xxxxxxxxxx> wrote:
> On Fri, Jul 1, 2016 at 11:35 AM, Kenneth Waegeman
> <kenneth.waegeman@xxxxxxxx> wrote:
>> Hi all,
>>
>> While syncing a lot of files to cephfs, our mds cluster got haywire: the
>> mdss have a lot of segments behind on trimming: (58621/30)
>> Because of this the mds cluster gets degraded. RAM usage is about 50GB. The
>> mdses were respawning and replaying continiously, and I had to stop all
>> syncs , unmount all clients and increase the beacon_grace to keep the
>> cluster up .
>>
>> [root@mds03 ~]# ceph status
>> cluster 92bfcf0a-1d39-43b3-b60f-44f01b630e47
>> health HEALTH_WARN
>> mds0: Behind on trimming (58621/30)
>> monmap e1: 3 mons at
>> {mds01=10.141.16.1:6789/0,mds02=10.141.16.2:6789/0,mds03=10.141.16.3:6789/0}
>> election epoch 170, quorum 0,1,2 mds01,mds02,mds03
>> fsmap e78658: 1/1/1 up {0=mds03=up:active}, 2 up:standby
>> osdmap e19966: 156 osds: 156 up, 156 in
>> flags sortbitwise
>> pgmap v10213164: 4160 pgs, 4 pools, 253 TB data, 203 Mobjects
>> 357 TB used, 516 TB / 874 TB avail
>> 4151 active+clean
>> 5 active+clean+scrubbing
>> 4 active+clean+scrubbing+deep
>> client io 0 B/s rd, 0 B/s wr, 63 op/s rd, 844 op/s wr
>> cache io 68 op/s promote
>>
>>
>> Now it finally is up again, it is trimming very slowly (+-120 segments /
>> min)
>
> Hmm, so it sounds like something was wrong that got cleared by either
> the MDS restart or the client unmount, and now it's trimming at a
> healthier rate.
>
> What client (kernel or fuse, and version)?
>
> Can you confirm that the RADOS cluster itself was handling operations
> reasonably quickly? Is your metadata pool using the same drives as
> your data? Were the OSDs saturated with IO?
>
> While the cluster was accumulating untrimmed segments, did you also
> have a "client xyz failing to advanced oldest_tid" warning?
This does not prevent MDS from trimming log segment.
>
> It would be good to clarify whether the MDS was trimming slowly, or
> not at all. If you can reproduce this situation, get it to a "behind
> on trimming" state, and the stop the client IO (but leave it mounted).
> See if the (x/30) number stays the same. Then, does it start to
> decrease when you unmount the client? That would indicate a
> misbehaving client.
Behind on trimming on single MDS cluster should be caused by either
slow rados operations or MDS trim too few log segments on each tick.
Kenneth, could you try setting mds_log_max_expiring to a large value
(such as 200)
Regards
Yan, Zheng
>
> John
> _______________________________________________
> ceph-users mailing list
> ceph-users@xxxxxxxxxxxxxx
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
