Virtualization still takes too much of a toll, especially on i/o-intensive applications, to be considered for HPC. Message-passing applications are heavily dependent on network latency, which virtualization affects. Also, there is no way to virtualize low-latency adapters (like Infiniband).
There is however one scenario where virtualization could be applied: cycle stealing for non-mpi applications. When latency is not critical, we can imagine deploying a hypervisor on the servers, with a "day job" partition and a "compute node" partition. Resources would be allocated dynamically to those (e.g. on schedule), according to the workload requirements. If anybody is interested in this scenario, please drop us a line here.