As far as I can understand the mentioned improvement is targeted on a
polling mode for a MSSQL, there is no wonder for observing relatively
high hypervisor CPU consumption without any sign of same consumption
in a guest itself. Certain kinds of applications (PBXes,
Chrome/Chromium, seemingly MSSQL) may poll some resource very
frequently, causing wakeups in the host for guest process and
increasing overall running time of a guest process. I`ve heard of none
of the mitigations for this class of workload issues, though you can
experiment with guest timer settings if you want to. As we are talking
in the libvir-list, it is safe to suggest to use native HyperV host to
eliminate this problem completely. :)