This topic provides information to plan for balancing the workload
(filesets) among the metadata servers.
Each metadata server in a cluster, including the master
metadata server, is assigned a workload. The workload is the amount
of processing that is required to manage the locks, leases, and metadata on
behalf of the clients when they request access to data that resides in a fileset
in the global namespace. To the client, the fileset appears to be another
directory, but to the metadata server, a fileset is an amount of workload.
For example, a fileset that contains data that is frequently accessed by clients
will have a higher workload than a fileset that stores archived data.
During client setup, a client is given the address of
one of the metadata servers for initial contact and metadata server cluster
discovery. When the client issues a request to access data, it is automatically
directed to the appropriate metadata server to obtain the metadata and locks
required to access the data.
Workload-balance methods
There are two methods for
assigning fileset to metadata servers:
- You can choose to statically assign filesets to metadata servers.
The master metadata server will make the assignment if the assigned metadata
server is running. If it is not running, the master metadata server will temporarily
assign the fileset to another metadata server. Once the assigned metadata
server is running, the master metadata server will automatically reassign
the fileset back to its assigned metadata server.
- You can choose to have the SAN File System dynamically assign a
fileset to a metadata server. The master metadata server will choose a metadata
server to be assigned based on a workload distribution algorithm. This algorithm
evenly distributes the filesets among the metadata servers in the cluster.
Dynamic-fileset assignment works best when there are many filesets with uniform
workloads.
High-availability considerations
You need enough
spare workload capacity such that a metadata server can be taken offline and
its filesets can be distributed to other metadata servers without overloading
them. To ensure that there is enough spare workload capacity:
- Ensure that in an N-server cluster configuration, each metadata
server runs at most (N–1)/N percent of capacity. This way, if
one metadata server goes offline, the remaining N–1 metadata servers
do not run at 100%.
- Because you cannot control which metadata server is the master, reserve
spare capacity (such as 5%) on each metadata server for the master workload.
- If you use static-fileset assignments, reserve a spare metadata server
in the cluster to take on the workload of another metadata server that goes
offline and preserve the workload balance. A spare metadata server is
an idle metadata server that has no statically assigned filesets. It can take
on the workload of any metadata server without impacting the remaining metadata
servers in the cluster.