Re: [RFC PATCH v1 00/26] migration: File based migration with multifd and fixed-ram

[Peter Xu]

Hi, Claudio,

Thanks for the context.

On Mon, Apr 03, 2023 at 09:47:26AM +0200, Claudio Fontana wrote:
> Hi, not sure if what is asked here is context in terms of the previous
> upstream discussions or our specific requirement we are trying to bring
> upstream.
>
> In terms of the specific requirement we are trying to bring upstream, we
> need to get libvirt+QEMU VM save and restore functionality to be able to
> transfer VM sizes of ~30 GB (4/8 vcpus) in roughly 5 seconds.  When an
> event trigger happens, the VM needs to be quickly paused and saved to
> disk safely, including datasync, and another VM needs to be restored,
> also in ~5 secs.  For our specific requirement, the VM is never running
> when its data (mostly consisting of RAM) is saved.
>
> I understand that the need to handle also the "live" case comes from
> upstream discussions about solving the "general case", where someone
> might want to do this for "live" VMs, but if helpful I want to highlight
> that it is not part of the specific requirement we are trying to address,
> and for this specific case won't also in the future, as the whole point
> of the trigger is to replace the running VM with another VM, so it cannot
> be kept running.

>From what I read so far, that scenario suites exactly what live snapshot
would do with current QEMU - that at least should involve a snapshot on the
disks being used or I can't see how that can be live.  So it looks like a
separate request.

> The reason we are using "migrate" here likely stems from the fact that
> existing libvirt code currently uses QMP migrate to implement the save
> and restore commands.  And in my personal view, I think that reusing the
> existing building blocks (migration, multifd) would be preferable, to
> avoid having to maintain two separate ways to do the same thing.  That
> said, it could be done in a different way, if the performance can keep
> up. Just thinking of reducing the overall effort and also maintenance
> surface.

I would vaguely guess the performance can not only keep up but better than
what the current solution would provide, due to the possibility of (1)
batch handling of continuous guest pages, and (2) completely no dirty
tracking overhead.

For (2), it's not about wr-protect page faults or vmexits due to PML being
full (because vcpus will be stopped anyway..), it's about enabling the
dirty tracking (which already contains overhead, especially when huge pages
are enabled, to split huge pages in EPT pgtables) and all the bitmap
operations QEMU does during live migration even if the VM is not live.

IMHO reusing multifd may or may not be a good idea here, because it'll of
course also complicate multifd code, hence makes multifd harder to
maintain, while not in a good way, because as I mentioned I don't think it
can use much of what multifd provides.

I don't have a strong opinion on the impl (even though I do have a
preference..), but I think at least we should still check on two things:

  - Being crystal clear on the use case above, and double check whether "VM
    stop" should be the default operation at the start of the new cmd - we
    shouldn't assume the user will be aware of doing this, neither should
    we assume the user is aware of the performance implications.

  - Making sure the image layout is well defined, so:

    - It'll be extensible in the future, and,

    - If someone would like to refactor it to not use the migration thread
      model anymore, the image format, hopefully, can be easy to keep
      untouched so it can be compatible with the current approach.

Just my two cents. I think Juan should have the best grasp on this.

Thanks,

-- 
Peter Xu