Open Stack

running Icehouse + Neutron ML2/OVS and network names spaces.

Was running well unitl recently, most recent change was switching to
Ceph RBD for ephemeral storage on the hypervisors (and glance). I
suspect this of being relevant because it makes the instances launch
much more quickly.

I haven't classified the breaking point but launching 64 instances
deterministically breaks the metadata agent.

The service seems to be running on the controller, but is not
listening in the network namespace.  It seems to require restarting
both the dhcp-agent and the metadata agent  to get it to go again.

Even in debug mode I get no errors in the logs.

Anyone seen this?

-Jon