Re: [Qemu-devel] [RFC 0/3] qga: add guest-get-memory-info (for 2.5)

[Marc-André Lureau]

Hi

----- Original Message -----
> On Fri, Jul 31, 2015 at 07:36:46PM +0200, address@hidden wrote:
> > From: Marc-André Lureau <address@hidden>
> > 
> > This series implement a new qemu guest agent command to get memory
> > information from the guest. This is based on ovirt-guest-agent
> > "memory-stats" message.
> > 
> > I couldn't find documentation for the ovirt message, but the list of
> > fields are summarized in this test
> > https://github.com/oVirt/ovirt-guest-agent/blob/master/tests/message_validator.py#L137
> > and according to how they are populated in the code, I adapted it to
> > the following GuestMemoryInfo structure fields:
> > 
> >  - mem-total: Total usable RAM.
> >  - mem-free:  Total of RAM that can be used without having to swap contents
> >  to disk.
> >  - mem-cached: In-RAM cache.
> >  - swap-total: Total amount of swap space available.
> >  - swap-free: Amount of swap space that is currently unused.
> >  - swap-in: Number of pages swapped-in per second.
> >  - swap-out: Number of pages swapped-out per second.
> >  - pf-major: Number of major page fault per second.
> >  - pf-minor: Number of minor page fault per second.
> > 
> > Implemented on Linux and Win32 based on ovirt implementations.
> 
> Interesting, currently the libvirt virDomainGetMemoryStats() API is backed
> by data we obtain from the virtio-balloon driver via QEMU monitor. For Linux
> at least this provides equiv of your mem-total, mem-free, swap-in, swap-out
> pf-major and pf-minor. So it lacks mem-cached, swap-total and swap-free
> I'm unclear in the Windows balloon driver supports this data or not too.

Ah, I should have thought about it! However, many guests don't have a balloon 
driver. So it could still be useful as a fallback imho

> > Note: the "per second" value differ between Linux and Win32. On Linux,
> > the value is computed based on the average since the last query,
> > however on win32 this seems to be an instantaneous value (they have
> > spikes, but often at 0). I have asked for help on SF:
> > 
> > http://serverfault.com/questions/709943/windows-equivalent-of-linux-vmstat-pswpin-and-pgfault
> > 
> > Related to RFE:
> > https://bugzilla.redhat.com/show_bug.cgi?id=1101915
> 
> I wonder if we would be better off extending the balloon driver to fill
> in the gaps we have there vs this guest agent impl. With the balloon
> agent we set things up so that the guest device periodically pushes the
> updated stats to QEMU. So when we're querying QEMU for the stats we don't
> actually block waiting on the guest OS at all, QEMU can answer directly.
> This feels more appealing that querying the guest agent where we have
> no reasonable expectation of prompt response. With a large enough number
> of guests, I think the balloon driver push approach will scale better
> than a guest agent approach.

I see, I think it really depends on the use case. Clearly for debugging, it 
doesn't need to be push-based. For load-balancing, I don't know either.

Please Jiri give some details on how these fields are used by ovirt. Thanks