The Performance Penalty of Virtualization

If you’ve spent any time with virtualized environments you know how effective and productive they are. The process of expanding capacity for FaceDouble, for example, became significantly simpler once they moved to depolying virtual servers, and SmugMug has been singing the praises of Amazon’s EC2 with a clever system to provision and remove capacity based on load. My own experiments with Hadoop and EC2 have been similarly fruitful.

So I’m wondering what the downside to aggressively going virtual is - why not make all servers virtual?

The main issue that comes to mind is performance, or the loss thereof. Presumably the performance of a virtual server is less than that of the same server running directly on the native OS.

Just how much of a performance difference is there, say in terms of per request latency and capacity, for a web server, a database server, and a cpu-bound heavy computation server, for any of the common virtualization systems (Xen, VMWare, etc)? I haven’t seen any good materials on this, so if you have knowledge or pointers please let me know.