Infact, I was pushing you to do the performance analysis on how much they benefit in the VM workloads.

Do you missed VM configurations in this article? Like how many vcpus do you give to the guest, and did you run any other background workload VMs at the same time? Did you over commit the machine?

Also I notice that Raghu is doing some PLE optimization on KVM, did your take that into account?

What I was suggesting is running more benchmarks proving improvement of the patch on interesting workload themselves (like a big number of guests).
The benchmark I did was to prove that the patch was not introducing a performance regression in stock Linux, increasing, then, chances to see it upstreamed as soon as possible. But other aspects of this work can be further explored.

For a direct patch file, I think you should ask to Raghu directly.
Can I suggest to run a separate set of benchmarks rather than replicating this case?
Discussion has gone forward and it may be very interesting to try targeted benchmarks on the workload itself.