summaryrefslogtreecommitdiffstats
path: root/doc/cephfs/add-remove-mds.rst
diff options
context:
space:
mode:
Diffstat (limited to '')
-rw-r--r--doc/cephfs/add-remove-mds.rst106
1 files changed, 106 insertions, 0 deletions
diff --git a/doc/cephfs/add-remove-mds.rst b/doc/cephfs/add-remove-mds.rst
new file mode 100644
index 00000000..c695fbbb
--- /dev/null
+++ b/doc/cephfs/add-remove-mds.rst
@@ -0,0 +1,106 @@
+============================
+ Deploying Metadata Servers
+============================
+
+Each CephFS file system requires at least one MDS. The cluster operator will
+generally use their automated deployment tool to launch required MDS servers as
+needed. Rook and ansible (via the ceph-ansible playbooks) are recommended
+tools for doing this. For clarity, we also show the systemd commands here which
+may be run by the deployment technology if executed on bare-metal.
+
+See `MDS Config Reference`_ for details on configuring metadata servers.
+
+
+Provisioning Hardware for an MDS
+================================
+
+The present version of the MDS is single-threaded and CPU-bound for most
+activities, including responding to client requests. Even so, an MDS under the
+most aggressive client loads still uses about 2 to 3 CPU cores. This is due to
+the other miscellaneous upkeep threads working in tandem.
+
+Even so, it is recommended that an MDS server be well provisioned with an
+advanced CPU with sufficient cores. Development is on-going to make better use
+of available CPU cores in the MDS; it is expected in future versions of Ceph
+that the MDS server will improve performance by taking advantage of more cores.
+
+The other dimension to MDS performance is the available RAM for caching. The
+MDS necessarily manages a distributed and cooperative metadata cache among all
+clients and other active MDSs. Therefore it is essential to provide the MDS
+with sufficient RAM to enable faster metadata access and mutation.
+
+Generally, an MDS serving a large cluster of clients (1000 or more) will use at
+least 64GB of cache (see also :doc:`/cephfs/cache-size-limits`). An MDS with a larger
+cache is not well explored in the largest known community clusters; there may
+be diminishing returns where management of such a large cache negatively
+impacts performance in surprising ways. It would be best to do analysis with
+expected workloads to determine if provisioning more RAM is worthwhile.
+
+In a bare-metal cluster, the best practice is to over-provision hardware for
+the MDS server. Even if a single MDS daemon is unable to fully utilize the
+hardware, it may be desirable later on to start more active MDS daemons on the
+same node to fully utilize the available cores and memory. Additionally, it may
+become clear with workloads on the cluster that performance improves with
+multiple active MDS on the same node rather than over-provisioning a single
+MDS.
+
+Finally, be aware that CephFS is a highly-available file system by supporting
+standby MDS (see also :ref:`mds-standby`) for rapid failover. To get a real
+benefit from deploying standbys, it is usually necessary to distribute MDS
+daemons across at least two nodes in the cluster. Otherwise, a hardware failure
+on a single node may result in the file system becoming unavailable.
+
+Co-locating the MDS with other Ceph daemons (hyperconverged) is an effective
+and recommended way to accomplish this so long as all daemons are configured to
+use available hardware within certain limits. For the MDS, this generally
+means limiting its cache size.
+
+
+Adding an MDS
+=============
+
+#. Create an mds data point ``/var/lib/ceph/mds/ceph-${id}``. The daemon only uses this directory to store its keyring.
+
+#. Edit ``ceph.conf`` and add MDS section. ::
+
+ [mds.${id}]
+ host = {hostname}
+
+#. Create the authentication key, if you use CephX. ::
+
+ $ sudo ceph auth get-or-create mds.${id} mon 'profile mds' mgr 'profile mds' mds 'allow *' osd 'allow *' > /var/lib/ceph/mds/ceph-${id}/keyring
+
+#. Start the service. ::
+
+ $ sudo systemctl start mds.${id}
+
+#. The status of the cluster should show: ::
+
+ mds: ${id}:1 {0=${id}=up:active} 2 up:standby
+
+Removing an MDS
+===============
+
+If you have a metadata server in your cluster that you'd like to remove, you may use
+the following method.
+
+#. (Optionally:) Create a new replacement Metadata Server. If there are no
+ replacement MDS to take over once the MDS is removed, the file system will
+ become unavailable to clients. If that is not desirable, consider adding a
+ metadata server before tearing down the metadata server you would like to
+ take offline.
+
+#. Stop the MDS to be removed. ::
+
+ $ sudo systemctl stop mds.${id}
+
+ The MDS will automatically notify the Ceph monitors that it is going down.
+ This enables the monitors to perform instantaneous failover to an available
+ standby, if one exists. It is unnecessary to use administrative commands to
+ effect this failover, e.g. through the use of ``ceph mds fail mds.${id}``.
+
+#. Remove the ``/var/lib/ceph/mds/ceph-${id}`` directory on the MDS. ::
+
+ $ sudo rm -rf /var/lib/ceph/mds/ceph-${id}
+
+.. _MDS Config Reference: ../mds-config-ref