For the openstack part: We’re morphing Enterprise Openstack to be Carrier-grade Openstack (quote from Peter)


Spanning networks

Redhat RHEV can currently span networks between RHEV clusters and openstack clusters by sharing a neutron instance.

  • MOC v1: using multiple nova regions that share a keystone backend and a single neutron.
  • MOC v2: Some sort of horizon, non-trusting protocol that can span providers.


Neutron/GRE has a single point of failure

Distributing the neutron management would be key to a large open source deployment. Possible solutions:

  • Creating a bridge in KVM on each of the nodes that would allow un-proxied Internet access
  • Creating a non-neutron-managed VM that’s connected between the private and the VLAN that’s public
  • All external traffic (using the software router) travels through a single node.
  • S3/lustre/other guest-based network traffic will have to travel through that single router
  • Manila is a project to export a filesystem to a guest VM from the host (thus: could use the compute node’s networking to get around neutron limitation)
  • Peer to peer traffic is only coordinated centrally; packets flow between peers
  • May not be an issue for 10s of nodes
  • Even with the Neutron Node in active/passive mode for reliability, if the main node goes down, the cluster dies as the routes are lost
  • Openstack does not currently let you mix modes between Neutron (GRE) and Nova-network (VLAN-based) within a single cluster.

Hardware as a networking solution

  • VxLANs supports 24 bits of network (similar to VLANs)


Plan is to use glustre.


Can use CloudForms for billing. (Redhat product, but plans are to opensource it).

Upgrades/Running the latest

Openstack does not have a smooth upgrade path that preserves configuration and VMs.

Managing openstack

  • MOC, if we wanted to run on the latest, will need to help refine the Continuous Integration to define known-good snapshots between the various components.
  • Example: Currently, ping tests are done in CI, but bandwidth tests are not (a neutron patch recently broke compatibility).
  • MOC could have a testbed known to be experimental. After a week, it could be declared good.