Re: [Qemu-devel] [PATCH v7 0/5] virtio-balloon: free page hint reporting support

[Peter Xu <peterx@redhat.com>]

On Fri, Jun 01, 2018 at 03:21:54PM +0800, Wei Wang wrote:
> On 06/01/2018 12:58 PM, Peter Xu wrote:
> > On Tue, Apr 24, 2018 at 02:13:43PM +0800, Wei Wang wrote:
> > > This is the deivce part implementation to add a new feature,
> > > VIRTIO_BALLOON_F_FREE_PAGE_HINT to the virtio-balloon device. The device
> > > receives the guest free page hints from the driver and clears the
> > > corresponding bits in the dirty bitmap, so that those free pages are
> > > not transferred by the migration thread to the destination.
> > > 
> > > - Test Environment
> > >      Host: Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz
> > >      Guest: 8G RAM, 4 vCPU
> > >      Migration setup: migrate_set_speed 100G, migrate_set_downtime 2 second
> > > 
> > > - Test Results
> > >      - Idle Guest Live Migration Time (results are averaged over 10 runs):
> > >          - Optimization v.s. Legacy = 271ms vs 1769ms --> ~86% reduction
> > >      - Guest with Linux Compilation Workload (make bzImage -j4):
> > >          - Live Migration Time (average)
> > >            Optimization v.s. Legacy = 1265ms v.s. 2634ms --> ~51% reduction
> > >          - Linux Compilation Time
> > >            Optimization v.s. Legacy = 4min56s v.s. 5min3s
> > >            --> no obvious difference
> > > 
> > > - Source Code
> > >      - QEMU:  https://github.com/wei-w-wang/qemu-free-page-lm.git
> > >      - Linux: https://github.com/wei-w-wang/linux-free-page-lm.git
> > Hi, Wei,
> > 
> > I have a very high-level question to the series.
> 
> Hi Peter,
> 
> Thanks for joining the discussion :)

Thanks for letting me know this thread.  It's an interesting idea. :)

> 
> > 
> > IIUC the core idea for this series is that we can avoid sending some
> > of the pages if we know that we don't need to send them.  I think this
> > is based on the fact that on the destination side all the pages are by
> > default zero after they are malloced.  While before this series, IIUC
> > any migration will send every single page to destination, no matter
> > whether it's zeroed or not.  So I'm uncertain about whether this will
> > affect the received bitmap on the destination side.  Say, before this
> > series, the received bitmap will directly cover the whole RAM bitmap
> > after migration is finished, now it's won't.  Will there be any side
> > effect?  I don't see obvious issue now, but just raise this question
> > up.
> 
> This feature currently only supports pre-copy (I think the received bitmap
> is something matters to post copy only).
> That's why we have
> rs->free_page_support = ..&& !migrate_postcopy();

Okay.

> 
> > Meanwhile, this reminds me about a more funny idea: whether we can
> > just avoid sending the zero pages directly from QEMU's perspective.
> > In other words, can we just do nothing if save_zero_page() detected
> > that the page is zero (I guess the is_zero_range() can be fast too,
> > but I don't know exactly how fast it is)?  And how that would be
> > differed from this page hinting way in either performance and other
> > aspects.
> 
> I guess you referred to the zero page optimization. I think the major
> overhead comes to the zero page checking - lots of memory accesses, which
> also waste memory bandwidth. Please see the results attached in the cover
> letter. The legacy case already includes the zero page optimization.

I replied in the other thread.  We can discuss there altogether.

Actually after a second thought I think maybe what I worried there is
exactly the reason why we must send the zero page flag - otherwise
there can be stale non-zero page on destination.  Here "zero page" and
"freed page" is totally different idea since even if a page is zeroed
it might still be in use (not freed)!  While instead for a "free page"
even if it's non-zero we might be able to not send it at all, though I
am not sure whether that mismatch of data might cause any side effect
too. I think the corresponding question would be: if a page is freed
in Linux kernel, would its data matter any more?

Thanks,

-- 
Peter Xu