Re: [RFC] Performance monitoring units and KVM

[Markus Armbruster <armbru@redhat.com>]

Balaji Rao <balajirrao@gmail.com> writes:

> Hi all!
>
> Earlier it was suggested that we go ahead with emulating Perf Mon Events in 
> exposing it to the guest. The serious limitation in this approach is that we 
> end up exposing only a small number of events to the guest, even though the 
> host hardware is capable of much more. The only benefit this approach offers is 
> that, it doesn't break live migration.
>
> The other option is to pass through the real PMU to the guest. I believe this 
> approach is far better in the sense that,

Do we really have an either/or alternative here?

> 1. All the available events in the host hardware can be passed on to the guest, 
> which can be used by oprofile to profile the guest and trackdown slowdowns 
> introduced due to virtualization.
>
> 2. Its much cleaner and easier to pass through the PMU.
>
> Yes, this approach breaks live migration. Migration should not be possible 
> *only* when the PMU is being used by oprofile. We can mark the guest as 
> unmigratable in such situations. Once the PMU is not being used, migration can 
> be performed normally.
>
> Note, this requires a small change to oprofile source. Upon migration, oprofile 
> should be made to re-identify the CPU and use the perf mon events appropriate 
> to that CPU. I think this could be done by having a migrate_notifier, or 
> something like that..
>
> Please provide comments on this.

Different implementations of the same processor architecture have
different PMUs.  Existing software using the PMU (directly) knows
exactly what PMU to expect with a particular CPU.

If we want to run such software in a guest (say VTune under Windows),
we need to provide a virtual CPU that is sufficiently complete,
including the PMU.  This will be *costly* on most CPUs.  Bad vmexit
latencies.

Sometimes it doesn't matter when profiling slows down your system, as
long as the profile is still sufficiently accurate, e.g. when a
developer examines a program in a test bed.  At other times, such
overhead is simply unacceptable, e.g. when you examine a real system
in the field, to figure out why it misbehaves.

There are ways to use the PMU in guests that don't require costly
virtualization of the real PMU.  They put the guest's performance
monitoring interface at a level higher than hardware PMU.

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/