Re: [PATCH kernel 0/2] speed up live migration by skipping free pages

From: Dr. David Alan Gilbert
Date: Mon Apr 25 2016 - 09:17:29 EST


* Michael S. Tsirkin (mst@xxxxxxxxxx) wrote:
> On Mon, Apr 25, 2016 at 05:38:30PM +0530, Amit Shah wrote:
> > On (Mon) 25 Apr 2016 [14:04:06], Michael S. Tsirkin wrote:
> > > On Mon, Apr 25, 2016 at 11:36:41AM +0530, Amit Shah wrote:
> > > > On (Tue) 19 Apr 2016 [22:34:32], Liang Li wrote:
> > > > > Current QEMU live migration implementation mark all guest's RAM pages
> > > > > as dirtied in the ram bulk stage, all these pages will be processed
> > > > > and it consumes quite a lot of CPU cycles and network bandwidth.
> > > > >
> > > > > From guest's point of view, it doesn't care about the content in free
> > > > > page. We can make use of this fact and skip processing the free
> > > > > pages, this can save a lot CPU cycles and reduce the network traffic
> > > > > significantly while speed up the live migration process obviously.
> > > > >
> > > > > This patch set is the kernel side implementation.
> > > > >
> > > > > The virtio-balloon driver is extended to send the free page bitmap
> > > > > from guest to QEMU.
> > > > >
> > > > > After getting the free page bitmap, QEMU can use it to filter out
> > > > > guest's free pages. This make the live migration process much more
> > > > > efficient.
> > > > >
> > > > > In order to skip more free pages, we add an interface to let the user
> > > > > decide whether dropping the cache in guest during live migration.
> > > >
> > > > So if virtio-balloon is the way to go (i.e. speed is acceptable), I
> > > > just have one point then. My main concern with using (or not using)
> > > > virtio-balloon was that a guest admin is going to disable the
> > > > virtio-balloon driver entirely because the admin won't want the guest
> > > > to give away pages to the host, esp. when the guest is to be a
> > > > high-performant one.
> > >
> > > The result will be the reverse of high-performance.
> > >
> > > If you don't want to inflate a balloon, don't.
> > >
> > > If you do but guest doesn't respond to inflate requests,
> > > it's quite reasonable for host to kill it -
> > > there is no way to distinguish between that and
> > > guest being malicious.
> >
> > With the new command I'm suggesting, the guest will let the host know
> > that it has enabled this option, and it won't free up any RAM for the
> > host.
> >
> > Also, just because a guest doesn't release some memory (which the
> > guest owns anyway) doesn't make it malicious, and killing such guests
> > is never going to end well for that hosting provider.
> >
> > > I don't know of management tools doing that but
> > > it's rather reasonable. What does happen is
> > > some random guest memory is pushed it out to swap,
> > > which is likely much worse than dropping unused memory
> > > by moving it into the balloon.
> >
> > Even if the host (admin) gave a guarantee that there won't be any
> > ballooning activity involved that will slow down the guest, a guest
> > admin can be paranoid enough to disable ballooning. If, however, this
> > is made known to the host, it's likely a win-win situation because the
> > host knows the guest needs its RAM, and the guest can still use the
> > driver to send stats which the host can use during migration for
> > speedups.
> >
> >
> > Amit
>
> We'd need to understand the usecase better to design a good interface
> for this. AFAIK the normal usecase for ballooning is for
> memory overcommit: asking guest to free up memory might work
> better than swap which makes host initiate a bunch of IO.
> How is not inflating in this case a good idea?
> I'm afraid I don't understand why was inflating balloon
> requested if we do not want the guest to inflate the balloon. What does
> "paranoid" mean in this context? This seems to imply some kind of
> security concern. Is guest likely to need all of its memory or a
> specific portion of it? Is it likely to be a static configuration or a
> dynamic one? If dynamic, does guest also want to avoid deflating the
> balloon or only inflating it?

The semantics we want here are subtly different from ballooning;

a) In ballooning the host is asking for the guest to give up some ram;
that's not quite the case here - although if the guest dropped some
caches that *might* be helpful.
b) In ballooning we're interested in just amounts of free memory; here
we care about _which_ pages are free.
c) We don't want to make it hard for the guest to start using some of the
memory again; if it was hard then we suddenly get back into a balancing
act of needing to quantatively talk about how much free RAM we have,
and worry about whether we shouldn't tell the host about a small pot
we keep back etc.

Dave

> --
> MST
--
Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK