On Wed, 26 Oct 2016, Trygve Vea wrote: > ----- Den 26.okt.2016 14:41 skrev Sage Weil [email protected]: > > On Wed, 26 Oct 2016, Trygve Vea wrote: > >> Hi, > >> > >> We have two Ceph-clusters, one exposing pools both for RGW and RBD > >> (OpenStack/KVM) pools - and one only for RBD. > >> > >> After upgrading both to Jewel, we have seen a significantly increased CPU > >> footprint on the OSDs that are a part of the cluster which includes RGW. > >> > >> This graph illustrates this: http://i.imgur.com/Z81LW5Y.png > > > > That looks pretty significant! > > > > This doesn't ring any bells--I don't think it's something we've seen. Can > > you do a 'perf top -p `pidof ceph-osd`' on one of the OSDs and grab a > > snapshot of the output? It would be nice to compare to hammer but I > > expect you've long since upgraded all of the OSDs... > > # perf record -p 18001 > ^C[ perf record: Woken up 57 times to write data ] > [ perf record: Captured and wrote 18.239 MB perf.data (408850 samples) ] > > > This is a screenshot of one of the osds during high utilization: > http://i.imgur.com/031MyIJ.png
It looks like a ton of time spent in std::string methods and a lot more map<sring,ghobject_t> than I would expect. Can you do a perf record -p `pidof ceph-osd` -g perf report --stdout > Link to download binary format sent directly to you. > > > Your expectation about upgrades is correct. We actually had some > problems performing the upgrade, so we ended up re-initializing the osds > as empty and backfill into jewel. When we first started them on jewel, > they ended up blocking Hrm, this is a new one for me too. They've all been upgraded now? It would be nice to see a log or backtrace to see why they got stuck. Thanks! sage > I want to add that the resource usage isn't flat - this is a day graph > of one of the osd servers: http://i.imgur.com/MLfoVgE.png > > > > Regards > -- > Trygve Vea > _______________________________________________ > ceph-users mailing list > [email protected] > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
