Re: [Qemu-devel] [PATCH qemu v4 18/18] memory: Give memory_region_transaction_commit a hint

[Paolo Bonzini]

----- Original Message -----
> From: "Alexey Kardashevskiy" <address@hidden>
> To: "Paolo Bonzini" <address@hidden>, address@hidden
> Sent: Thursday, September 21, 2017 2:02:02 AM
> Subject: Re: [PATCH qemu v4 18/18] memory: Give 
> memory_region_transaction_commit a hint
> 
> On 21/09/17 03:14, Paolo Bonzini wrote:
> > On 20/09/2017 13:46, Alexey Kardashevskiy wrote:
> >> This extends memory_region_transaction_commit() to receive a MR as
> >> if it is a root MR or its topmost parent is, then we can only rebuild
> >> its FlatView and update it for address spaces sharing it.
> >>
> >> The optimization gets disabled though if there is full update about to
> >> commit.
> >>
> >> memory_region_set_enabled() is a special case here, it does not use
> >> a hint when MR is being disabled.
> >>
> >> On POWER8 with 255 CPUs, 255 virtio-net, 40 PCI bridges guest this brings
> >> down the boot time from 20s to 12s, the total memory footprint
> >> goes down (17G -> 8G).
> > 
> > I think this is incorrect if MR has an alias (no matter if enabling or
> > disabling)?
> 
> Hmmm. Right. I can add a aliases_nr counter to an MR (it does not even have
> to go down as aliases are not creared/destroyed often) and take the slow
> path if it is not zero, does it make sense?

It should be a separate series anyway.  Let's get the infrastructure in
so it's easier to experiment.

Paolo