Jon Stokes over at Ars Techncia has a really good article on how virtualization is a prime culprit for resource contention on multi core processors.

Jon sheds light on just a few of the challenges vendors like Intel have to address to make heavy duty Virtualization/Consolidation ’sustainable’:  Intel’s vision of the future of the “processor” is a heterogeneous network on a chip (NoC). This NoC may contain cores of different types, some of which are highly application-specific. These cores will run threads that have very different resource needs and usage patterns than the more general-purpose x86 cores

Another snippet from his article ‘Cache and memory in the many-core era‘ is below:

Virtualization has the potential to exacerbate the resource contention problem tremendously, since different workloads running in different virtual machines—workloads managed by OS instances that are mutually unaware of one another—can create a kind of “perfect storm” and pull overall system performance down through the floor.

Or, to put this in plainer English by using an analogy, the VM contention problem is kind of like a two-family condo where both floors are on the same (aging) breaker box. On a Friday night in the dead of summer, if the undergrad girls in the top unit are running eight lights, two air conditioners, four curling irons, and an entertainment system, and the couple down below is running an air conditioner, a washer and dryer, and a dishwasher, then something’s bound to blow. Within each unit, the inhabitants may be carefully monitoring their electricity usage so as not to overtax the aging circuits, but the lack of cross-unit coordination is a recipe for periodic blackouts.

(To unpack the analogy briefly, the two condo units would be two virtual machines, and the person keeping track of electricity usage in each unit would be the OS.)

Virtualization-related resource contention is already a real problem, and there have been a number of attempts by Intel and others in the academic community to profile how different types of workloads interact (c.f. “Virtualization in the Enterprise,” Intel Technology Journal, v10, issue 3). What these studies have found is that degenerate cases, where resource contention causes performance degradation across all the VMs in a system, are not uncommon on enterprise workloads.

I hate to trivialise the article, but it’s interesting and recommended reading.