Using HP Serviceguard for Linux with VMware virtual machines - Technical white paper

Cluster configuration options

A Serviceguard cluster that includes virtual machine nodes can consist of:

• Virtual machines on the same host (cluster in a box; not recommended)

• Virtual machines on separate hosts

• Virtual machine and physical nodes

• All of the above

If a cluster is configured with multiple virtual machines running on the same host, together with virtual machines

running on other hosts or physical servers, you need to be aware of the possibility of data corruption if an application

fails over between virtual machines running on the same ESX host. If one guest node that is part of the cluster hangs and

a package fails over to another guest node on the same host, there is a very small possibility that I/O pending for the

first guest does not complete before the package is started on the second guest.

Unsafe configuration: Because of the risk of data corruption, a cluster consisting entirely of multiple VMs running on

the same ESX host is not HA safe and should be avoided in a production environment.

Figure 11. HP Serviceguard cluster of virtual machine guest and physical node

Safe configuration: Serviceguard takes care of data integrity when applications fail over between

Virtual machines running on separate physical hosts. This means that you can safely configure a cluster in which

multiple virtual machines are running on one host, provided packages are configured in such a way that the failover does

not happen between two virtual machines running on the same ESX server.

A Serviceguard cluster consisting of a virtual machine guest and a physical server is shown in figure 11. This cluster

provides HA for the applications against failures of physical nodes, VMware ESX hypervisor, VM guest, and failure of the

application itself. A failed application can be restarted on the same VM guest, or failed over to the physical node. This

configuration allows consolidation of multiple active-standby clusters where the primary active node remains a physical

node and multiple standby nodes are consolidated as virtual machine guests on one host.