If you eliminate the hypervisor, you don't have vCPUs at all, so I'm not sure how this is a useful comparison.
The question of IPC performance is an interesting one. We've been building up a database of open source results that show wildly diverging results on different architectures. Surf through http://fable.io for it, or read this work in progress:
http://anil.recoil.org/papers/drafts/2012-usenix-ipc-draft1.... http://anil.recoil.org/talks/fosdem-io-2012.pdf (fosdem slides)
The TL;DR of these numbers is that it's very hard to make firm performance hypotheses about IPC across architectures, NUMA and hypervisors.
Are you assuming that vCPUs are scheduled or pinned? It wasn't clear from skimming the paper. I would agree that two levels of scheduling are bad but if that's the problem pinning should fix it.
The situation may change if you eliminate the hypervisor. My intuition is that a single multi-threaded process with work-stealing will be faster than separate processes or VMs due to better load balance (see PX vs. 1X dynos) and faster inter-thread communication (if your app has any communication).