Le mercredi 06 mars 2013 à 09:39 +0100, Olivier Bonvalet a écrit : > > Le lundi 04 mars 2013 à 09:02 -0800, Gregory Farnum a écrit : > > On Mon, Mar 4, 2013 at 12:19 AM, Olivier Bonvalet <ceph.list@xxxxxxxxx> wrote: > > > Hello, > > > > > > I have a running cluster, which use the (previous) default CRUSH rule, > > > with : > > > step take default > > > step choose firstn 0 type osd > > > step emit > > > > > > Since I have multiple OSD in the same host, I need to change that to > > > have better redondancy. > > > My crush map use types datacenter → room → network → rack → host → osd. > > > > > > For now I would like to replicate per "network", I suppose I can simply > > > replace "step choose firstn 0 type osd" per "step choose firstn 0 type > > > network" ; but is it safe to do that on a running cluster ? > > > > Not quite — you'll want to do "step chooseleaf firstn 0 type network". > > See http://ceph.com/docs/master/rados/operations/crush-map for details > > on this stuff. > > Yes, it is safe to do on a running cluster. > > > > > > > And I suppose it will throw a lot of data move, is it possible to do > > > that «slowly», to avoid production slowdown ? > > > > Yeah, it's likely to move the entire contents. If you're on Argonaut > > you're kind of stuck; if Bobtail then you have some control through > > the "osd_max_backfills" option (defaults to 10) that specifies how > > many moves each OSD will let itself be involved in at once. You should > > practice on a non-production cluster first to make sure you don't run > > into issues with pgtemp or OSDMap generation, though. > > -Greg > > > > Thanks for your answer. So I made some tests on a dedicated spool, I was > able to move data from «platter» to «SSD» very well, it's great. > > But I can't obtain that per "network" neither per "host" setup : > with 2 hosts, each one with 2 OSD, and with a pool with use only 1 > replica (so, 2 copies), I tried this rule : > > rule rbdperhost { > ruleset 5 > type replicated > min_size 1 > max_size 10 > step take default > step chooseleaf firstn 0 type host > step emit > } > > > As a result I obtain some PG which stuck in «active+remapped» state. > When querying one of this PG, I see that CRUSH find only one OSD up for > this one and can't find an other OSD to set replica. > > If I well understand, in this case the "chooseleaf firstn 0 type host" > say to Ceph to choose 2 differents hosts, then in each of them choose > one OSD. So with 2 hosts, it should works, no ? > > Thanks, > Olivier B. > > So, as said on IRC, it's solved. My rules were not working, and after use of «tunables» it's ok. I love that feature of changing data spreading on live ! _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |