Configuring multiple active MDS daemons

Also known as: multi-mds, active-active MDS

Each CephFS file system is configured for a single active MDS daemon by default. To scale metadata performance for large scale systems, you may enable multiple active MDS daemons, which will share the metadata workload with one another.

When should I use multiple active MDS daemons?

You should configure multiple active MDS daemons when your metadata performance is bottlenecked on the single MDS that runs by default.

Adding more daemons may not increase performance on all workloads. Typically, a single application running on a single client will not benefit from an increased number of MDS daemons unless the application is doing a lot of metadata operations in parallel.

Workloads that typically benefit from a larger number of active MDS daemons are those with many clients, perhaps working on many separate directories.

Increasing the MDS active cluster size

Each CephFS file system has a max_mds setting, which controls how many ranks will be created. The actual number of ranks in the file system will only be increased if a spare daemon is available to take on the new rank. For example, if there is only one MDS daemon running, and max_mds is set to two, no second rank will be created. (Note that such a configuration is not Highly Available (HA) because no standby is available to take over for a failed rank. The cluster will complain via health warnings when configured this way.)

Set max_mds to the desired number of ranks. In the following examples the “fsmap” line of “ceph status” is shown to illustrate the expected result of commands.

# fsmap e5: 1/1/1 up {0=a=up:active}, 2 up:standby

ceph fs set <fs_name> max_mds 2

# fsmap e8: 2/2/2 up {0=a=up:active,1=c=up:creating}, 1 up:standby
# fsmap e9: 2/2/2 up {0=a=up:active,1=c=up:active}, 1 up:standby

The newly created rank (1) will pass through the ‘creating’ state and then enter this ‘active state’.

Standby daemons

Even with multiple active MDS daemons, a highly available system still requires standby daemons to take over if any of the servers running an active daemon fail.

Consequently, the practical maximum of max_mds for highly available systems is at most one less than the total number of MDS servers in your system.

To remain available in the event of multiple server failures, increase the number of standby daemons in the system to match the number of server failures you wish to withstand.

Decreasing the number of ranks

Reducing the number of ranks is as simple as reducing max_mds:

# fsmap e9: 2/2/2 up {0=a=up:active,1=c=up:active}, 1 up:standby
ceph fs set <fs_name> max_mds 1
# fsmap e10: 2/2/1 up {0=a=up:active,1=c=up:stopping}, 1 up:standby
# fsmap e10: 2/2/1 up {0=a=up:active,1=c=up:stopping}, 1 up:standby
# fsmap e10: 1/1/1 up {0=a=up:active}, 2 up:standby

The cluster will automatically stop extra ranks incrementally until max_mds is reached.

See CephFS Administrative commands for more details which forms <role> can take.

Note: stopped ranks will first enter the stopping state for a period of time while it hands off its share of the metadata to the remaining active daemons. This phase can take from seconds to minutes. If the MDS appears to be stuck in the stopping state then that should be investigated as a possible bug.

If an MDS daemon crashes or is killed while in the up:stopping state, a standby will take over and the cluster monitors will against try to stop the daemon.

When a daemon finishes stopping, it will respawn itself and go back to being a standby.

Manually pinning directory trees to a particular rank

In multiple active metadata server configurations, a balancer runs which works to spread metadata load evenly across the cluster. This usually works well enough for most users but sometimes it is desirable to override the dynamic balancer with explicit mappings of metadata to particular ranks. This can allow the administrator or users to evenly spread application load or limit impact of users’ metadata requests on the entire cluster.

The mechanism provided for this purpose is called an export pin, an extended attribute of directories. The name of this extended attribute is Users can set this attribute using standard commands:

setfattr -n -v 2 path/to/dir

The value of the extended attribute is the rank to assign the directory subtree to. A default value of -1 indicates the directory is not pinned.

A directory’s export pin is inherited from its closest parent with a set export pin. In this way, setting the export pin on a directory affects all of its children. However, the parents pin can be overridden by setting the child directory’s export pin. For example:

mkdir -p a/b
# "a" and "a/b" both start without an export pin set
setfattr -n -v 1 a/
# a and b are now pinned to rank 1
setfattr -n -v 0 a/b
# a/b is now pinned to rank 0 and a/ and the rest of its children are still pinned to rank 1

Setting subtree partitioning policies

It is also possible to setup automatic static partitioning of subtrees via a set of policies. In CephFS, this automatic static partitioning is referred to as ephemeral pinning. Any directory (inode) which is ephemerally pinned will be automatically assigned to a particular rank according to a consistent hash of its inode number. The set of all ephemerally pinned directories should be uniformly distributed across all ranks.

Ephemerally pinned directories are so named because the pin may not persist once the directory inode is dropped from cache. However, an MDS failover does not affect the ephemeral nature of the pinned directory. The MDS records what subtrees are ephemerally pinned in its journal so MDS failovers do not drop this information.

A directory is either ephemerally pinned or not. Which rank it is pinned to is derived from its inode number and a consistent hash. This means that ephemerally pinned directories are somewhat evenly spread across the MDS cluster. The consistent hash also minimizes redistribution when the MDS cluster grows or shrinks. So, growing an MDS cluster may automatically increase your metadata throughput with no other administrative intervention.

Presently, there are two types of ephemeral pinning:

Distributed Ephemeral Pins: This policy indicates that all of a directory’s immediate children should be ephemerally pinned. The canonical example would be the /home directory: we want every user’s home directory to be spread across the entire MDS cluster. This can be set via:

setfattr -n -v 1 /cephfs/home

Random Ephemeral Pins: This policy indicates any descendent sub-directory may be ephemerally pinned. This is set through the extended attribute with the value set to the percentage of directories that should be pinned. For example:

setfattr -n -v 0.5 /cephfs/tmp

Would cause any directory loaded into cache or created under /tmp to be ephemerally pinned 50 percent of the time.

It is recomended to only set this to small values, like .001 or 0.1%. Having too many subtrees may degrade performance. For this reason, the config mds_export_ephemeral_random_max enforces a cap on the maximum of this percentage (default: .01). The MDS returns EINVAL when attempting to set a value beyond this config.

Both random and distributed ephemeral pin policies are off by default in Octopus. The features may be enabled via the mds_export_ephemeral_random and mds_export_ephemeral_distributed configuration options.

Ephemeral pins may override parent export pins and vice versa. What determines which policy is followed is the rule of the closest parent: if a closer parent directory has a conflicting policy, use that one instead. For example:

mkdir -p foo/bar1/baz foo/bar2
setfattr -n -v 0 foo
setfattr -n -v 1 foo/bar1

The foo/bar1/baz directory will be ephemerally pinned because the foo/bar1 policy overrides the export pin on foo. The foo/bar2 directory will obey the pin on foo normally.

For the reverse situation:

mkdir -p home/{patrick,john}
setfattr -n -v 1 home
setfattr -n -v 2 home/patrick

The home/patrick directory and its children will be pinned to rank 2 because its export pin overrides the policy on home.

If a directory has an export pin and an ephemeral pin policy, the export pin applies to the directory itself and the policy to its children. So:

mkdir -p home/{patrick,john}
setfattr -n -v 0 home
setfattr -n -v 1 home

The home directory inode (and all of its directory fragments) will always be located on rank 0. All children including home/patrick and home/john will be ephemerally pinned according to the distributed policy. This may only matter for some obscure performance advantages. All the same, it’s mentioned here so the override policy is clear.