We have 12 osds per host, so we've gone conservative and set recovery max active to 1 and max backfills to 4. We also set nodown prior to adding a new osd since we saw flapping can be even more problematic in recovery. On Apr 12, 2013 8:04 PM, "Dave Spano" <dsp...@optogenics.com> wrote:
> What are you settings for recovery max active and backfill? Just curious. > > Dave Spano > > > ------------------------------ > *From: *"Erdem Agaoglu" <erdem.agao...@gmail.com> > *To: *"Dave Spano" <dsp...@optogenics.com> > *Cc: *"Stefan Priebe - Profihost AG" <s.pri...@profihost.ag>, > "ceph-users" <ceph-users@lists.ceph.com> > *Sent: *Friday, April 12, 2013 12:48:05 PM > *Subject: *Re: [ceph-users] ceph recovering results in offline VMs > > We are also seeing a similar problem which we believe it's #3737. Our VMs > (running mongodbs) were being completely frozen for 2-3 minutes (sometimes > longer) while adding a new OSD. We have reduced recovery max active and > backfill settings and ensured that we have RBD caching and now it seems > things are better. We still see some increase in iowaits but VM's continue > to function. > > But that i guess depends on what VM actually does at that moment. We did > some fio tests before running actual services and what we saw was that > while individual read or write tests were able to survive OSD addition with > some degraded performance, concurrent read-write tests (rw and randrw in > fio talk) were completely stalled. I mean the VM was able to function in > individual read or write tests even if performance sometimes drops to 0 > iops, but it was frozen in rw/randrw test in addition to dropping to 0 iops. > > BTW Stefan, i'm in no way experienced with ceph and i don't know about > your OSD's but 8128 pgs for a 8TB cluster seems too much. Or is it OK when > disks are SSDs? > > > On Fri, Apr 12, 2013 at 5:23 PM, Dave Spano <dsp...@optogenics.com> wrote: > >> Very interesting. I ran into the same thing yesterday when I added SATA >> disks to the cluster. I was about to return them for SAS drives instead >> because of how long it took, and how slow some of my RBDs got. >> >> Are most people using SATA 7200 RPM drives? My concern was with Oracle >> DBs. Postgres doesn't seem to have as much of a problem running on an RBD, >> but I noticed a marked difference with Oracle. >> >> Dave Spano >> >> >> >> ------------------------------ >> *From: *"Stefan Priebe - Profihost AG" <s.pri...@profihost.ag> >> *To: *"Wido den Hollander" <w...@42on.com> >> *Cc: *ceph-users@lists.ceph.com >> *Sent: *Wednesday, April 10, 2013 3:51:23 PM >> *Subject: *Re: [ceph-users] ceph recovering results in offline VMs >> >> >> Am 10.04.2013 um 21:36 schrieb Wido den Hollander <w...@42on.com>: >> >> > On 04/10/2013 09:16 PM, Stefan Priebe wrote: >> >> Hello list, >> >> >> >> i'm using ceph 0.56.4 and i've to replace some drives. But while ceph >> is >> >> backfilling / recovering all VMs have high latencies and sometimes >> >> they're even offline. I just replace one drive at a time. >> >> >> >> I putted in the new drives and i'm reweighting them from 0.0 to 1.0 in >> >> 0.1 steps. >> >> >> >> I already lowered osd recovery max active = 2 and osd max backfills = >> 3, >> >> but when i put them back at 1.0 the vms are nearly all down. >> >> >> >> Right now some drives are SSDs so they're a lot faster than the HDDs >> i'm >> >> going to replace them too. >> >> >> >> Nothing in the logs but it is recovering at 3700MB/s that this is not >> >> possible on SATA HDDs is clear. >> >> >> >> Log example: >> >> 2013-04-10 20:55:33.711289 mon.0 [INF] pgmap v9293315: 8128 pgs: 233 >> >> active, 7876 active+clean, 19 active+recovery_wait; 557 GB data, 1168 >> GB >> >> used, 7003 GB / 8171 GB avail; 2108KB/s wr, 329op/s; 31/309692 degraded >> >> (0.010%); recovering 840 o/s, 3278MB/s >> > >> > There is a issue about this in the tracker, I saw it this week but I'm >> not able to find it anymore. >> >> 3737? >> >> > I'm seeing this as well, when the cluster is recovering RBD images tend >> to get very sluggish. >> > >> > Most of the time I'm blaiming the CPUs in the OSDs for it, but I've >> also seen it on faster systems. >> >> I've 3,6Ghz xeons with just 4 osds per host. >> >> Stefan >> _______________________________________________ >> ceph-users mailing list >> ceph-users@lists.ceph.com >> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> >> >> _______________________________________________ >> ceph-users mailing list >> ceph-users@lists.ceph.com >> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> >> > > > -- > erdem agaoglu > >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com