Re: [PATCH 0/3] migration: Downtime tracepoints

[Peter Xu]

On Thu, Oct 26, 2023 at 05:06:37PM +0100, Joao Martins wrote:
> On 26/10/2023 16:53, Peter Xu wrote:
> > This small series (actually only the last patch; first two are cleanups)
> > wants to improve ability of QEMU downtime analysis similarly to what Joao
> > used to propose here:
> > 
> >   https://lore.kernel.org/r/20230926161841.98464-1-joao.m.martins@oracle.com
> > 
> Thanks for following up on the idea; It's been hard to have enough bandwidth 
> for
> everything on the past set of weeks :(

Yeah, totally understdood.  I think our QE team pushed me towards some
series like this, while my plan was waiting for your new version. :)

Then when I started I decided to go into per-device.  I was thinking of
also persist that information, but then I remembered some ppc guest can
have ~40,000 vmstates..  and memory to maintain that may or may not regress
a ppc user.  So I figured I should first keep it simple with tracepoints.

> 
> > But with a few differences:
> > 
> >   - Nothing exported yet to qapi, all tracepoints so far
> > 
> >   - Instead of major checkpoints (stop, iterable, non-iterable, resume-rp),
> >     finer granule by providing downtime measurements for each vmstate (I
> >     made microsecond to be the unit to be accurate).  So far it seems
> >     iterable / non-iterable is the core of the problem, and I want to nail
> >     it to per-device.
> > 
> >   - Trace dest QEMU too
> > 
> > For the last bullet: consider the case where a device save() can be super
> > fast, while load() can actually be super slow.  Both of them will
> > contribute to the ultimate downtime, but not a simple summary: when src
> > QEMU is save()ing on device1, dst QEMU can be load()ing on device2.  So
> > they can run in parallel.  However the only way to figure all components of
> > the downtime is to record both.
> > 
> > Please have a look, thanks.
> >
> 
> I like your series, as it allows a user to pinpoint one particular bad device,
> while covering the load side too. The checkpoints of migration on the other 
> hand
> were useful -- while also a bit ugly -- for the sort of big picture of how
> downtime breaks down. Perhaps we could add that /also/ as tracepoitns without
> specifically commiting to be exposed in QAPI.
> 
> More fundamentally, how can one capture the 'stop' part? There's also time 
> spent
> there like e.g. quiescing/stopping vhost-net workers, or suspending the VF
> device. All likely as bad to those tracepoints pertaining device-state/ram
> related stuff (iterable and non-iterable portions).

Yeah that's a good point.  I didn't cover "stop" yet because I think it's
just more tricky and I didn't think it all through, yet.

The first question is, when stopping some backends, the vCPUs are still
running, so it's not 100% clear to me on which should be contributed as
part of real downtime.

Meanwhile that'll be another angle besides vmstates: need to keep some eye
on the state change handlers, and that can be a device, or something else.

Did you measure the stop process in some way before?  Do you have some
rough number or anything surprising you already observed?

Thanks,

-- 
Peter Xu