No Hyper-V enlightenments are passed to Windows guests by default, the default set lives in layered products. We need to do some performance benchmarking and come up with a 'recommendation' for them.
Currently, we have the following:
hv-relaxed - windows activates it automatically when hypervisor CPU flag is present.
hv-vapic - should speed things up in general (IRQs)
hv-time - should speed things up significantly
hv-crash - can we capture this info?
hv-reset - rather redundant
hv-vpindex - required for PV TLB flush/IPSs
hv-runtime - for stats
hv-synic - required for stimer
hv-stimer - speed things up, especially with the latest Win10 update
hv-frequencies - needed for Hyper-V on KVM (tsc page clocksource
hv-reenlightenment - also needed for Hyper-V on KVM (tsc page clocksource)
hv-tlbflush - speed things up in overcommited environments
We have another two in development: hv_vmcs and hv_ipi, we can discuss them when they're accepted upstream.
This is something we'll need to include in libosinfo and the templates
I have already asked for libosinfo and database schema to accomodate this kind of information.
Please note synic and stimer are necessary for good performance of the latest Windows update (https://bugzilla.redhat.com/show_bug.cgi?id=1610461) and will require libvirt/qemu/host kernel update.
Franecsco, I saw multiple fixes for hyperv getting merged, should this be in modified?
(In reply to Fabian Deutsch from comment #6)
> Franecsco, I saw multiple fixes for hyperv getting merged, should this be in
I think kubevirt core is fine, but for CNV we still need to update the templates and actually use the hyperv flags
evmcs is tracked in bug 1952551
Other than that, the following flags are set in Windows templates:
hv-time and hv-crash are not included in the templates; is this intentional?
@firstname.lastname@example.org hv-time and hv-crash are not included in the templates; should we include these features as well?
@email@example.com please attach these two flags as different stories to https://issues.redhat.com/browse/CNV-11966 we should understand their use-case before adding them to the templates
Once tracked on the epic, this bug should be good
Moving to verified.