Thin-Provisioning: Virtualization or Storage System?
Thin-provisioning, a great feature for virtual machines, uses the disk space required instead of what is allocated. I frequently refer to the Windows pie graph that displays used and free space as a good way to explain what thin-provisioned virtual disk files “cost” on disk. All major hypervisors support it, and there is somewhat of a debate as to whether or not the storage system or virtualization engine should manage thin-provisioning.
On the other hand, many storage systems in use for virtualization environment support a thin-provisioned volume. This is effectively aggregating the same benefit across many virtual machines. In the case of a volume that is presented to a physical server directly, thin-provisioned disks can be used in that way as well. Most mainstream storage products with decent management now support thin-provisioning of virtual machine storage types, primarily VMware's vStorage VMFS and Windows NTFS file systems. NTFS is probably marginally more supported simply due to Windows having broader support. A thin-provisioned volume on the storage system effectively is VMFS- or NTFS-aware for what is going on inside. This means that it knows if a .VMDK file is using its full allocation.
The prevailing thought is to have the storage system perform thin-provisioning for virtualization volumes. There is an non-quantified amount of overhead that may be associated with a write operation that would extend the file. You can hedge this off by using the 8 MB block size on your VMFS volumes, due to its built-in efficiencies or allocating a higher performance tier of disk. For the storage system to manage thin-provisioning, you're using on the disk controllers to manage that dynamic growth with direct access to write cache. What probably makes the least sense is to thin-provision on both environments. The disk savings would be minimal, yet overhead would be increased.
How do you approach thin-provisioning? Share your comments below.
Posted by Rick Vanover on 01/27/2010 at 12:47 PM