Open Stack

 Hello All,

 For the use case of attaching a volume created using cinder to an
 instance that is created in a compute node that represents a *cluster*
 requires
 
 1.       The volume to be presented to all the hosts in the cluster
 
 2.       The nova driver to be aware of the target LUN information as
 seen for each host and make the volume attach to the instance.
 
  
  In today's implementation of nova and cinder interfaces for attaching a
 volume using the /nova volume-attach/, the control/data flow between the
 two components is done based on the assumption that the compute node
 consists of only one host. However, some changes need to be done to
 handle clusters represented as a compute node. The problem was brought
 up for discussion in the Havana summit with the Cinder team and they
 were of the opinion that the problem can be resolved by the following
 approach
 
 1.       The nova manager will call the driver's get_volume_connector
 method to obtain the connector information consisting of the compute
 node's HBA information.
 
 2.       The driver that manages the cluster will now return the details
 of all the hosts in the cluster. This is returned in an additional key
 of the dict called clustered_hosts. The value will have the connector
 information for each host.
 
 3.       /[CHANGE proposed]/The nova manager will now be changed to be
 aware of the new key clustered_hosts and then iterate through each one
 and call the cinder APIs initialize_connection. Cinder drivers response
 to this will remain as it is today, i.e present the volume to the host
 and return the target information for the volume.
 
 4.       The nova manager will pass the collected target information
 back to the nova driver to attach the volume to the instance as a new
 key. The driver will now be aware of the new key and perform the volume
 attach to the instance.
 
  
 
 Let me know if the above approach is acceptable? Please add the
 necessary nova core team members for reviewing this approach so that we
 can discuss and finalize on the approach.
 

 The blueprint link that has additional details. Here is the link:
 https://blueprints.launchpad.net/cinder/+spec/multi-attach-volume
  
 
 The initial approach that was discussed in the Havana summit was with
 cinder doing the looping for each host (steps below). However, cinder
 team favored the solution where there would be minimal/no change in
 cinder but yet achieve the same result with the existing interfaces/data.
   
 
 1.       The nova manager will call the driver's get_volume_connector
 method to obtain the connector information consisting of the compute
 node's HBA information.
 
 2.       The driver that manages the cluster will now return the details
 of all the hosts in the cluster. This is returned in an additional key
 of the dict called clustered_hosts. The value will have the connector
 information for each host.
 
 3.       The nova manager will call the cinder APIs initialize_connection.
 
 4.       /[CHANGE proposed initially]/Cinder drivers will be aware of
 the new key clustered_hosts and cinder will perform the loop for each
 host and
 
 a.       present the volume to the host and collect the target
 information for the volume.
 
 b.      Collect target information of all hosts and send it back to nova
 in a new key
 
 5.       The nova manager will pass the collected target information
 back to the nova driver to attach the volume to the instance as a new
 key. The driver will now be aware of the new key and perform the volume
 attach to the instance.
 
  
 Thanks,
 Kiran