Hello, thanks for the info.
I'll give a try tomorrow. On one of my test I got the messages that yo say (wrongfully marked), but i've lowered other options and now is fine. For now the OSD are not reporting down messages even with an high load test, but I'll see the logs tomorrow to confirm.
The most of time the server is used as RO and the load is not high, so if an OSD is marked as down for some seconds is not a big problem (at least I think that recovery traffic is low because it only has to check that pgs are in both OSD).
Greetings and thanks again!
2017-06-15 18:13 GMT+02:00 David Turner <drakonstein@xxxxxxxxx>:
osd_heartbeat_grace is a setting for how many seconds since the last time an osd received a successful response from another osd before telling the mons that it's down. This is one you may want to lower from its default value of 20 seconds.mon_osd_min_down_reporters is a setting for how many osds need to report an osd as down before the mons will mark it as down. I recommend setting this to N+1 where N is how many osds you have in a node or failure domain. If you end up with a network problem and you have 1 osd node that can talk to the mons, but not the other osd nodes, then you will end up with that one node marking the entire cluster down while the rest of the cluster marks that node down. If your min_down_reporters is N+1, then 1 node cannot mark down the rest of the cluster. The default setting is 1 so that small test clusters can mark down osds, but if you have 3+ nodes, you should set it to N+1 if you can. Setting it to more than 2 nodes is equally as problematic. However, if you just want things to report as fast as possible, leaving this at 1 still might be optimal to getting it marked down sooner.The downside to lowering these settings is if OSDs are getting marked down for running slower, then they will re-assert themselves to the mons and end up causing backfilling and peering for no really good reason. You'll want to monitor your cluster for OSDs being marked down for a few seconds before marking themselves back up. You can see this in the OSD logs where the OSD says it was wrongfully marked down in one line and then the next is where it tells the mons it is actually up.On Thu, Jun 15, 2017 at 10:44 AM Daniel Carrasco <d.carrasco@xxxxxxxxx> wrote:I forgot to say that after upgrade the machine RAM to 4Gb, the OSD daemons has started to use only a 5% (about 200MB). Is like magic, and now I've about 3.2Gb of free RAM.Greetings!!______________________________2017-06-15 15:08 GMT+02:00 Daniel Carrasco <d.carrasco@xxxxxxxxx>:Finally, the problem was W3Total Cache, that seems to be unable to manage HA and when the master redis host is down, it stop working without try the slave.I've added some options to make it faster to detect a down OSD and the page is online again in about 40s.[global]fsid = Hiddenmon_initial_members = alantra_fs-01, alantra_fs-02, alantra_fs-03mon_host = 10.20.1.109,10.20.1.97,10.20.1.216 auth_cluster_required = cephxauth_service_required = cephxauth_client_required = cephxosd mon heartbeat interval = 5osd mon report interval max = 10mon osd report timeout = 15osd fast fail on connection refused = Truepublic network = 10.20.1.0/24osd pool default size = 2Greetings and thanks for all your help.2017-06-14 23:09 GMT+02:00 David Turner <drakonstein@xxxxxxxxx>:I've used the kernel client and the ceph-fuse driver for mapping the cephfs volume. I didn't notice any network hiccups while failing over, but I was reading large files during my tests (and live) and some caching may have hidden hidden network hiccups for my use case.Going back to the memory potentially being a problem. Ceph has a tendency to start using 2-3x more memory while it's in a degraded state as opposed to when everything is health_ok. Always plan for over-provisioning your memory to account for a minimum of 2x. I've seen clusters stuck in an OOM killer death spiral because it kept killing OSDs for running out of memory, that caused more peering and backfilling, ... which caused more OSDs to be killed by OOM killer.On Wed, Jun 14, 2017 at 5:01 PM Daniel Carrasco <d.carrasco@xxxxxxxxx> wrote:Is strange because on my test cluster (three nodes) with two nodes with OSD, and all with MON and MDS, I've configured the size to 2 and min_size to 1, I've restarted all nodes one by one and the client loose the connection for about 5 seconds until connect to other MDS.Are you using ceph client or kernel client?I forgot to say that I'm using Debian 8.Anyway, maybe the problem was what I've said before, the clients connection with that node started to fail, but the node was not officially down. And it wasn't a client problem, because it happened on both clients and on my monitoring service at same time.Just now I'm not on the office, so I can't post the config file. Tomorrow I'll send it.Anyway, is the basic file generated by ceph-deploy with client network and min_size configurations. Just like my test config.Thanks!!, and greetings!!______________________________El 14 jun. 2017 10:38 p. m., "David Turner" <drakonstein@xxxxxxxxx> escribió:I have 3 ceph nodes, size 3, min_size 2, and I can restart them all 1 at a time to do ceph and kernel upgrades. The VM's running out of ceph, the clients accessing MDS, etc all keep working fine without any problem during these restarts. What is your full ceph configuration? There must be something not quite right in there.On Wed, Jun 14, 2017 at 4:26 PM Daniel Carrasco <d.carrasco@xxxxxxxxx> wrote:El 14 jun. 2017 10:08 p. m., "David Turner" <drakonstein@xxxxxxxxx> escribió:Not just the min_size of your cephfs data pool, but also your cephfs_metadata pool.Both were at 1. I don't know why because I don't remember to have changed the min_size and the cluster has 3 odd from beginning (I did it on another cluster for testing purposes, but I don't remember to have changed on this). I've changed both to two, but after the fail.About the size, I use 50Gb because it's for a single webpage and I don't need more space.I'll try to increase the memory to 3Gb.Greetings!!On Wed, Jun 14, 2017 at 4:07 PM David Turner <drakonstein@xxxxxxxxx> wrote:Ceph recommends 1GB of RAM for ever 1TB of OSD space. Your 2GB nodes are definitely on the low end. 50GB OSDs... I don't know what that will require, but where you're running the mon and mds on the same node, I'd still say that 2GB is low. The Ceph OSD daemon using 1GB of RAM is not surprising, even at that size.When you say you increased the size of the pools to 3, what did you do to the min_size? Is that still set to 2?On Wed, Jun 14, 2017 at 3:17 PM Daniel Carrasco <d.carrasco@xxxxxxxxx> wrote:Finally I've created three nodes, I've increased the size of pools to 3 and I've created 3 MDS (active, standby, standby).Today the server has decided to fail and I've noticed that failover is not working... The ceph -s command shows like everything was OK but the clients weren't able to connect and I had to restart the failing node and reconect the clients manually to make it work again (even I think that the active MDS was in another node).I don't know if maybe is because the server was not fully down, and only some connections were failing. I'll do some tests too see.Another question: How many memory needs a node to work?, because I've nodes with 2GB of RAM (one MDS, one MON and one OSD), and they have an high memory usage (more than 1GB on the OSD).The OSD size is 50GB and the data that contains is less than 3GB.Thanks, and Greetings!!______________________________2017-06-12 23:33 GMT+02:00 Mazzystr <mazzystr@xxxxxxxxx>:Since your app is an Apache / php app is it possible for you to reconfigure the app to use S3 module rather than a posix open file()? Then with Ceph drop CephFS and configure Civetweb S3 gateway? You can have "active-active" endpoints with round robin dns or F5 or something. You would also have to repopulate objects into the rados pools.Also increase that size parameter to 3. ;-)Lots of work for active-active but the whole stack will be much more resilient coming from some with a ClearCase / NFS / stale file handles up the wazoo backgroundOn Mon, Jun 12, 2017 at 10:41 AM, Daniel Carrasco <d.carrasco@xxxxxxxxx> wrote:2017-06-12 16:10 GMT+02:00 David Turner <drakonstein@xxxxxxxxx>:I have an incredibly light-weight cephfs configuration. I set up an MDS on each mon (3 total), and have 9TB of data in cephfs. This data only has 1 client that reads a few files at a time. I haven't noticed any downtime when it fails over to a standby MDS. So it definitely depends on your workload as to how a failover will affect your environment.On Mon, Jun 12, 2017 at 9:59 AM John Petrini <jpetrini@xxxxxxxxxxxx> wrote:______________________________We use the following in our ceph.conf for MDS failover. We're running one active and one standby. Last time it failed over there was about 2 minutes of downtime before the mounts started responding again but it did recover gracefully.[mds]max_mds = 1mds_standby_for_rank = 0mds_standby_replay = true___
John Petrini
_________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph. com
Thanks to both.Just now i'm working on that because I needs a very fast failover. For now the tests give me a very fast response when an OSD fails (about 5 seconds), but a very slow response when the main MDS fails (I've not tested the real time, but was not working for a long time). Maybe was because I created the other MDS after mount, because I've done some test just before send this email and now looks very fast (i've not noticed the downtime).Greetings!!--_________________________________________ Daniel Carrasco Marín
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph. com
--_________________________________________ Daniel Carrasco Marín
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________ _________________
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
--_________________________________________ Daniel Carrasco Marín
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________ --_________________________________________ Daniel Carrasco Marín
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________ _________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph. com
_________________________________________
Daniel Carrasco Marín
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________
Ingeniería para la Innovación i2TIC, S.L.
Tlf: +34 911 12 32 84 Ext: 223
www.i2tic.com
_________________________________________
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
