May I assume that you regard a VM with an Exchange server as one of the more important ones ?
If yes then from a recovery point of View there is only one good option:
- create three vmdks as Eager zeroed thick provisioned
- if possible create the vmdks while the datastore is still empty
- in case you can use an empty datastore create the largest vmdks first - (the idea is to enable ESXi to allocate all vmdks in a single piece)
It was just yesterday that I had to tell a customer : "your Exchange VM can not be recovered because you used thin-provisioning"
For performance I vote for:
scsi0:0.filename = "c-drive.vmdk"
scsi1:0.filename = "d-drive.vmdk"
scsi2:0.filename = "e-drive.vmdk"
Note the use of 3 scsi-controllers !
I always thought that this would make no big difference until we made comparison tests - the differences were not drastical but we got between 5% and 10% better overall performance using a dedicated scsi-controller for each vmdk.
We only compared 3 single vmdks using a single controller versus 3 vmdks with 3 controllers.
Personally I would only use vmdks with more than one partition for experimental VMs.
Only exception is the 100 MB boot partition you get with Window 7 and later - they are acceptable because they do not affect the performance.
