[Linux-ha-dev] Xen resource live migration and memory management

[Sebastian Reitenbach]

Hi,

right now, as far as I can see, the Xen live migration doesn't work together 
with Xen.

The problem described in short words:
- with xen memory management enabled, the available memory on all hosts is 
distributed equally on all running domU's on a physical node.

- with live migration, there must be enough resources available on the node 
where the domU will migrate to.

- as there is all memory used, due to the memory management, the live 
migration will fail.

- a resource script on one node, e.g. the Xen script, is not really able to 
manipulate sth. on other hosts in the cluster

To solve the problem I thought about the following:

1. create a XenMigrationHelperResource, that can do the following:
  - figure out the amount of running domU's
  - figure out the amount of maximum available memory per host
  - post both values into the CIB, using attrd_updater
  - resize memory on the domU's running on the same host as where the  
XenMigrationHelperResource runs on

2. using this resource, in e.g. a two node cluster:
 - create a group, ordered but not collocated
 - first add the  XenMigrationHelperResource
 - then the corresponding Xen resource

add a colocation constraint with -INFINITY, to keep the Xen resource always on 
a different host than the XenMigrationHelperResource

3. migrate the Xen resource, due to the order, the  XenMigrationHelperResource 
starts its work:
  - count running xen domains, and figure out maximum available memory, and 
updated the CIB with attrd_updater
 - shrink the memory on the running domU's equally, to make memory available 
for the new domU to be migrated to that host
- then it migrates away on the other host

- then the Xen resource starts migration, it will read the values for the 
target node posted via attrd_updater, and it will shrink the memory, so that 
it will fit on the other machine.
- then it will start the migration of the domU, which now has good chances to 
end successfully
- at the end io migration, the Xen resource will raise the memory on the host 
where it migrated from

Is this something that could work out, or maybe I think too complicated, and I 
overlooked sth. to solve the problem much more easier?

any input is highly appreciated.

regards
Sebastian