Today I did some experimenting to see what was possible. I configured a clustering scenario which is in fact not supported by VMware. I created a Microsoft Cluster pair (nodeA and nodeB) using virtual disks only which were all stored on the SAN. NodeA ran on ESX1 and NodeA ran ESX2. In turn ESX1 and ESX2 where part of DRS/HA Cluster.
What’s not supported by such a configuration. Well, to get VMware Support the boot disks of VM Cluster must be on local storage. Knowing this fact, I went ahead and ignored this configuration. What did I find?
Well, DRS does nothing in the area of VMotion. VMotion and VM clustering are hard-wired incompatible. You can’t manually VMotion VM cluster full-stop. Basically, VMware must check the configuration of the VM to see if the SCSI1 controller are set to be shared either virtually or physically. This is a requirement for the VM cluster to work. However, DRS was able to use “Initial Placement” to decide the best ESX host to power on NodeA and NodeB. Additionally, anti-affinity rules were obeyed which were configured to separate NodeA and NodeB.
Experiences with VMware HA were more interesting. I was able to crash the ESX1 host, and as a consequence NodeA crashed as well. Microsoft Clustering handled the failover to NodeB. In a short-while my other ESX hosts noticed that ESX1 had failed, and VMware HA fired up nodeA on one my remain hosts.
So technically, VMware HA and VM Clustering can be made to work together – but only if you are prepared to breach the supported configuration. I wouldn’t recommend doing this. I work in a Test & Dev environment where such experiments are tolerated. But it wouldn’t be a configuration I would feel comfortable with in a production environment.