Literature says that performance for guest is near-native, and that's
what I wanna test. And if somebody has already done this kind of test,
I could know what to expect.
No one else has likely tested your proprietary application. You should test without expectation.
I have, un-expectantly, seen an application in a DomU outperform baremetal in some corner cases with memory-access-intensive applications. I think this had to do with the DomU (paravirt) being put entirely into HugePages, where baremetal was using standard 4K pages, thus increasing TLB efficiency and reducing memory accesses.
The performance impact can vary largely based on workload, up or down, believe it or not.
-JR