Presentation was by David Nalley.

CloudStack is multi-tenant:

• Built around user abstraction (can’t see hypervisor/storage/etc)
• Networking seperation – every acct has at least one dedicated & isolated VLAN
• Hardware can be dedicated to users (for compliance, etc)

Networking:

• Broken down by how they are isolated, and if hardware is physical or virtual
• Offered Services: DHCP, VLAN allocation, Firewall, NAT/Port forwarding, Routing, VPN, Load Balancing
• Can also manage Big IPs, NetScaler, Juniper SRX, Cisco Nexus, Nixera, Brocade load balancers

Security groups:

• Traditional isolation has been by VLANs., but issues
  • Hard limit of 4096 VLANs
  • Supporting HW is expensive
• Security groups – layer 3 isolates. Up to 50K bridged devices managed
• Assumes flat layer 2 network, no multicast. Only will have hypervisors on the layer 2 network

High Availability

• Means services stay up – not just bringing them back up
• CloudStack is not high availablity, but its a tool to increase availability

Allocation algorithms

• How VMs, storage, etc are allocated
  • First fit: First place that it comes across that it fits
  • Fill first: fill up the resource before moving on
  • Disperse: (per zone)
  • Create your own
• Resources can be tagged, things will only get deployed to a tagged resource if they request the tag
• OS preference – for socket licensing, etc

Usage

• Not billing, but stats to bill against

Architecture

• Secondary storage: snapshots, templates for VMs,
• Historically NFS – option of object storage (swift, caringo, glusterfs)
• Managed by secondary storage vm, always at least 1 per zone, can scale out.
• Primary storage – NSF, iSCSI, and CLVM. (NSF/iSCSI don’t scale well.)
• Shared mountpoint – anything all hypervisors can mount and write to
• if apps can be architected to be stateless, its fast its local, etc

Resource dvision

• Zones
  • Specific geographic location, shares secondary storage resource, single network model
• Pods
  • Rack or row of racks, shares guest network
• Clusters
  • max of 8-15 machins, homogenity (same hupervisor, same COUs, same networking)

Networks

• Mgt nw
• Private nw
• Public nw
• guest nw

Mgt server – stateless (state is in DB), all UI functionality is an API call.

last part was a diagram of the environment. Pretty cool