Another now-false assumption is about the nature of OS images. In the physical world, the ratio of OS instances to boxes was 1. VM technology blows this up -- not only with all the live VMs, but all the other images that are suspended, saved for things like backup or rollback. It is not hard to build a datacenter with tens, even hundreds, of thousands of VM images. That sucks up a lot of disk space.
The irony is that the vast majority of the bits inside those images are exactly the same, because the VMs are cloned from a small set of golden masters, and a lot of the bits in a VM don't change as the VM runs. De-duplication could help here, but in this use case, de-dupe is more of a bug than a feature. If the storage infrastructure was built right, dupes wouldn't be created in the first place.
A couple other assumptions vitiated by virtual servers are that storage provisioning is a relatively infrequent activity, and that clustering of servers is rare and usually small scale. In the VM world, every creation of a VM requires provisioning of storage, and to take advantage of features like live migration across physical servers, we want all our servers to be in one big cluster.
One of the assumptions most aggravating to IT managers relates to the nature of hardware. In the old physical world, serious server applications were run on big, proprietary, expensive servers. It made sense for storage to match. Now, servers are small, non-proprietary, and spectacularly inexpensive. In stark contrast, storage systems are eerily reminiscent of the mainframe and minicomputer of yesteryear. The assumption that storage should be proprietary, should tie advanced software features to a particular brand of hardware, and be super expensive doesn't make sense.
InfoWorld: How do these false assumptions cause problems for people running virtual servers?
Virsto: The first symptom is that, all things being equal, virtual servers consume 15 to 25 percent more storage than physical servers. VM sprawl has a direct hardware budget impact, as well as an ongoing operating cost burden.
Second is the VM I/O blender phenomenon. Take a server that can do 1,000 I/O operations per second (IOPS) with one OS running. Now start eight VMs running the benchmark simultaneously. Aggregate IOPS can fall by 80 percent. Across all eight VMs, you're getting around 200 IOPS, or about 25 IOPS per VM. For good reason, IT architects are loath to virtualize I/O intensive applications.
Third, IT operations people find that basic storage management gets much more complex. Old backup techniques break down in a production virtual server environment. Provisioning complexity and performance management become burdensome. The hypervisor has many benefits, but simplifying storage management isn't one of them.