Description of problem: ---------------------- The ceph balancer commands are slow to respond or do not respond at all in our scale lab env (510 OSDs, 927TB). In our small, simple (no OSP) cluster (20 OSDs, 2.7 TB), the balancer status and eval commands return immediately but in the larger env (Ceph 3.1 in OSP13) the status command has taken upwards of 30sec while the eval command on the smallest (empty) pool does not return. Component Version-Release: ------------------------- Red Hat Enterprise Linux Server release 7.5 (Maipo) 3.10.0-862.11.6.el7.x86_64 ceph-common.x86_64 2:12.2.4-42.el7 @rhos-13.0-signed libcephfs2.x86_64 2:12.2.4-42.el7 @rhos-13.0-signed openstack-cinder.noarch 1:12.0.3-2.el7ost @rhos-13.0-signed openstack-dashboard.noarch 1:13.0.1-2.el7ost @rhos-13.0-signed openstack-dashboard-theme.noarch 1:13.0.0-1.el7ost @rhos-13.0-signed openstack-ec2-api.noarch 6.0.1-0.20180329214259.219dbc7.el7ost openstack-glance.noarch 1:16.0.1-3.el7ost @rhos-13.0-signed openstack-heat-agents.noarch 1.5.4-0.20180308153305.ecf43c7.el7ost openstack-heat-api.noarch 1:10.0.1-2.el7ost @rhos-13.0-signed openstack-heat-api-cfn.noarch 1:10.0.1-2.el7ost @rhos-13.0-signed openstack-heat-common.noarch 1:10.0.1-2.el7ost @rhos-13.0-signed openstack-heat-engine.noarch 1:10.0.1-2.el7ost @rhos-13.0-signed openstack-ironic-api.noarch 1:10.1.3-5.el7ost @rhos-13.0-signed openstack-ironic-common.noarch 1:10.1.3-5.el7ost @rhos-13.0-signed openstack-keystone.noarch 1:13.0.1-1.el7ost @rhos-13.0-signed openstack-manila.noarch 1:6.0.1-2.el7ost @rhos-13.0-signed openstack-manila-share.noarch 1:6.0.1-2.el7ost @rhos-13.0-signed openstack-manila-ui.noarch 2.13.0-5.el7ost @rhos-13.0-signed openstack-mistral-api.noarch 6.0.3-1.el7ost @rhos-13.0-signed openstack-mistral-common.noarch 6.0.3-1.el7ost @rhos-13.0-signed openstack-mistral-engine.noarch 6.0.3-1.el7ost @rhos-13.0-signed openstack-neutron.noarch 1:12.0.3-2.el7ost @rhos-13.0-signed openstack-neutron-common.noarch 1:12.0.3-2.el7ost @rhos-13.0-signed openstack-neutron-ml2.noarch 1:12.0.3-2.el7ost @rhos-13.0-signed openstack-nova-api.noarch 1:17.0.5-3.d7864fbgit.el7ost openstack-nova-common.noarch 1:17.0.5-3.d7864fbgit.el7ost openstack-nova-scheduler.noarch 1:17.0.5-3.d7864fbgit.el7ost openstack-panko-api.noarch 4.0.2-1.el7ost @rhos-13.0-signed openstack-panko-common.noarch 4.0.2-1.el7ost @rhos-13.0-signed openstack-sahara.noarch 1:8.0.1-2.el7ost @rhos-13.0-signed openstack-sahara-ui.noarch 8.0.1-1.el7ost @rhos-13.0-signed openstack-selinux.noarch 0.8.14-14.el7ost @rhos-13.0-signed openstack-swift-account.noarch 2.17.1-0.20180314165245.caeeb54.el7ost openstack-swift-container.noarch 2.17.1-0.20180314165245.caeeb54.el7ost puppet-ceph.noarch 2.5.0-1.el7ost @rhos-13.0-signed How reproducible: ---------------- consistent Steps to Reproduce: ------------------ 1. verify ceph-mgr is running ... systemctl status -l ceph-mgr@overcloud-controller-0 2. ceph mgr module enable balancer 3. time ceph balancer status 4. time ceph balancer eval backups Actual results: -------------- In the larger env the balancer 'status' responded in 0m29.255s while the 'eval' of the backups pool hung. Expected results: ---------------- The expected output of the balancer commands. Additional info: --------------- I disabled the balancer, restarted ceph-mgr, and re-enabled the balancer and then the status command responded almost immediately. The eval command does not return whether I pass a pool name or not. An strace of the PID shows it talking to OSDs looking for connection but gets ECONNREFUSED.
Tim, can you set me up with access to the cluster so I can take a closer look?
Raising priority of this problem because now we know that ceph-mgr can become effectively disabled by this problem. So ceph-mgr was bouncing continuously. This was hard to watch as docker logs would not show what was happening, but eventually I removed the CGroup limit of 1 GB from the systemd unit file /etc/systemd/system/ceph-mgr@.service and did systemctl daemon-reload and restarted the service and now it seems to be running. Suggest that we remove this CGroup limit or make it much much larger, because it can cause the ceph-mgr to get a memory allocation failure and crash, which impacts the entire ceph cluster. For example, "ceph pg dump" no longer works.
what version of ceph-ansible were you using as it generates the unit files?
(In reply to John Fulton from comment #6) > what version of ceph-ansible were you using as it generates the unit files? I'm afraid the scale lab allocation ended yesterday and was immediately wiped so I can not provide the exact version used. I saved off the undercloud ~stack homedir but I can't find that version buried in a log anywhere. I can say that the OC was director deployed OSP13 w/Ceph 3.0. I slipped the Ceph 3.1 (ceph tag: 3-13) container image in later to test newer code validating the OSD memory usage fix. The ceph-ansible version would have been whatever rhos-release gave me for '13' at the time.
A possibly related bz that could provide an alternative explanation for why the ceph-mgr balancer module failed is: https://bugzilla.redhat.com/show_bug.cgi?id=1593110#c9 We found out today that ceph-mgr CGroup limit was set by ceph-ansible. However, in latest available RHCS 3.1: http://download-node-02.eng.bos.redhat.com/rel-eng/RHCEPH-3.1-RHEL-7-20180927.1/compose/Tools/x86_64/os/Packages/ceph-ansible-3.1.5-1.el7cp.noarch.rpm The default ceph mgr docker memory limit is 1 GB, as shown here: root@perfshift02:~/bene/usr/share/ceph-ansible # grep -ir docker_memory_limit . ,,, ./group_vars/mgrs.yml.sample:#ceph_mgr_docker_memory_limit: 1g ... ./roles/ceph-mgr/defaults/main.yml:ceph_mgr_docker_memory_limit: 1g ./roles/ceph-mgr/templates/ceph-mgr.service.j2: --memory={{ ceph_mgr_docker_memory_limit }} \ ... This is unchanged in http://download.eng.bos.redhat.com/composes/auto/ceph-3.2-rhel-7/latest-RHCEPH-3-RHEL-7/compose/Tools/x86_64/os/Packages/ceph-ansible-3.2.0-0.1.beta3.el7cp.noarch.rpm
targeting RHCS 3.2
confirmed mon to 3g and osd to 5g as per: https://bugzilla.redhat.com/show_bug.cgi?id=1591871 but still 1g for mgrs: https://github.com/ceph/ceph-ansible/blob/824ec6d256fc23794d69dd82f789fb05ef5c7bb6/roles/ceph-mgr/defaults/main.yml#L27
Oops wrong BZ sorry
Since the problem described in this bug report should be resolved in a recent advisory, it has been closed with a resolution of ERRATA. For information on the advisory, and where to find the updated files, follow the link below. If the solution does not work for you, open a new bug report. https://access.redhat.com/errata/RHBA-2019:0020