CES address failover and distribution policies

When a Cluster Export Services (CES) node leaves the GPFS cluster, any CES IP addresses that are assigned to that node are moved to CES nodes still within the cluster. Additionally, certain error conditions and administrative operations can cause a node to release its addresses to be reassigned to other nodes.

As CES nodes enter and leave the GPFS cluster, the addresses are distributed among the nodes according to the address distribution policy that is selected. In addition, you can disable automatic address distribution to allow the user to manually maintain the address-to-node assignments.

The address distribution policy is set with the mmces command:
mmces address policy [even-coverage | balanced-load | node-affinity | none]
The following list describes each type of address distribution policy:
even-coverage
Distributes the addresses among the available nodes. The even-coverage policy is the default address distribution policy.
Note: If you have multiple CES networks, even IP address distribution in each network for every node might not be considered. The overall number of IP addresses on each node or CES group takes precedence.
Specify mmces address move to manually move IP addresses from one node to another node.
balanced-load
Distributes the addresses to approach an optimized load distribution. The loads (network and CPU) on all the nodes are monitored. Addresses are moved based on given policies for optimized load throughout the cluster.
Attention: In some cases, it is possible that balanced-load does not create an optimal load distribution. For robust balancing, use even-coverage load distribution that balances the number of CES-IPs as evenly as possible.
node-affinity
Attempts to keep an address on the node to which the user manually assigned it. If the mmces address add command is used with the --ces-node option, the address is marked as being associated with that node. Similarly, if an address is moved with the mmces address move command, the address is marked as being associated with the destination node. Any automatic movement, such as reassigning a down node's addresses, does not change this association. Addresses that are enabled with no node specification do not have a node association.
Addresses that are associated with a node but assigned to a different node are moved back to the associated node if possible.
Important: Run the following command to view the node affinity preferred nodes: 
mmces address list --full-list
If a CES IP address does not have the preferred affinity, you can move the CES IP address to the specified affinity node by running this command: 
mmces address move
If you manually move the IP address by specifying mmces address move, it sets the affinity of all moved CES IP addresses to the target node.
Automatic address distribution is performed in the background in a way as to not disrupt the protocol servers more than necessary. If you want immediate redistribution of the addresses, use the mmces command to force an immediate rebalance:
mmces address move --rebalance

In order to prevent an interruption in service, IP addresses that have attributes assigned to them (for example: object_database_node or object_singleton_node) are not rebalanced.

IPs without a node assignment will not fail back if they were re-assigned due to some failure condition. You can run the following command to assign a node to an existing CES IP address:
mmces address move --ces-ip {IP[,IP...]} --ces-node NODE
You can run the following command to check the current assignment:
mmces address list --full-list

You can further control the assignment of CES addresses by placing nodes or addresses in CES groups. For more information, see Configuring CES protocol service IP addresses.