Since BSD is heavy on sandbox models I'm assuming that will also be true here and they're not implementing a para-virtualized architecture. That being said, the host kernel will have to wait for NOOP on X amount of cores in order to permit the VM to run. You may also gain performance by disabling hyperthreading. With HT your cores share resources for each thread. This is great for small time stuff that may only ever need a fraction of time on the bus or make only a couple of changes to the cache. But compute heavy actions should have full access.