RE: [RFC qemu 0/4] A PV solution for live migration optimization
From: Li, Liang Z
Date: Mon Mar 14 2016 - 23:31:48 EST
> > > Hi,
> > > I'm just catching back up on this thread; so without reference to
> > > any particular previous mail in the thread.
> > >
> > > 1) How many of the free pages do we tell the host about?
> > > Your main change is telling the host about all the
> > > free pages.
> > Yes, all the guest's free pages.
> > > If we tell the host about all the free pages, then we might
> > > end up needing to allocate more pages and update the host
> > > with pages we now want to use; that would have to wait for the
> > > host to acknowledge that use of these pages, since if we don't
> > > wait for it then it might have skipped migrating a page we
> > > just started using (I don't understand how your series solves that).
> > > So the guest probably needs to keep some free pages - how many?
> > Actually, there is no need to care about whether the free pages will be
> used by the host.
> > We only care about some of the free pages we get reused by the guest,
> > The dirty page logging can be used to solve this, starting the dirty
> > page logging before getting the free pages informant from guest. Even
> > some of the free pages are modified by the guest during the process of
> > getting the free pages information, these modified pages will be traced by
> the dirty page logging mechanism. So in the following
> migration_bitmap_sync() function.
> > The pages in the free pages bitmap, but latter was modified, will be
> > reset to dirty. We won't omit any dirtied pages.
> > So, guest doesn't need to keep any free pages.
> OK, yes, that works; so we do:
> * enable dirty logging
> * ask guest for free pages
> * initialise the migration bitmap as everything-free
> * then later we do the normal sync-dirty bitmap stuff and it all just works.
> That's nice and simple.
> > > 2) Clearing out caches
> > > Does it make sense to clean caches? They're apparently useful data
> > > so if we clean them it's likely to slow the guest down; I guess
> > > they're also likely to be fairly static data - so at least fairly
> > > easy to migrate.
> > > The answer here partially depends on what you want from your
> > > if you're after the fastest possible migration time it might make
> > > sense to clean the caches and avoid migrating them; but that might
> > > be at the cost of more disruption to the guest - there's a trade off
> > > somewhere and it's not clear to me how you set that depending on
> > > guest/network/reqirements.
> > >
> > Yes, clean the caches is an option. Let the users decide using it or not.
> > > 3) Why is ballooning slow?
> > > You've got a figure of 5s to balloon on an 8GB VM - but an
> > > 8GB VM isn't huge; so I worry about how long it would take
> > > on a big VM. We need to understand why it's slow
> > > * is it due to the guest shuffling pages around?
> > > * is it due to the virtio-balloon protocol sending one page
> > > at a time?
> > > + Do balloon pages normally clump in physical memory
> > > - i.e. would a 'large balloon' message help
> > > - or do we need a bitmap because it tends not to clump?
> > >
> > I didn't do a comprehensive test. But I found most of the time
> > spending on allocating the pages and sending the PFNs to guest, I
> > don't know that's the most time consuming operation, allocating the pages
> or sending the PFNs.
> It might be a good idea to analyse it a bit more to convince people where the
> problem is.
Yes, I will try to measure the time spending on different parts.
> > > * is it due to the madvise on the host?
> > > If we were using the normal balloon messages, then we
> > > could, during migration, just route those to the migration
> > > code rather than bothering with the madvise.
> > > If they're clumping together we could just turn that into
> > > one big madvise; if they're not then would we benefit from
> > > a call that lets us madvise lots of areas?
> > >
> > My test showed madvise() is not the main reason for the long time,
> > only taken 10% of the total inflating balloon operation time.
> > Big madvise can more or less improve the performance.
> OK; 10% of the total is still pretty big even for your 8GB VM.
> > > 4) Speeding up the migration of those free pages
> > > You're using the bitmap to avoid migrating those free pages; HPe's
> > > patchset is reconstructing a bitmap from the balloon data; OK, so
> > > this all makes sense to avoid migrating them - I'd also been thinking
> > > of using pagemap to spot zero pages that would help find other zero'd
> > > pages, but perhaps ballooned is enough?
> > >
> > Could you describe your ideal with more details?
> At the moment the migration code spends a fair amount of time checking if a
> page is zero; I was thinking perhaps the qemu could just open
> /proc/self/pagemap and check if the page was mapped; that would seem
> cheap if we're checking big ranges; and that would find all the balloon pages.
Even if virtio-balloon is not enabled, it can be used to find the pages that never used
> > > 5) Second-migrate
> > > Given a VM where you've done all those tricks on, what happens when
> > > you migrate it a second time? I guess you're aiming for the guest
> > > to update it's bitmap; HPe's solution is to migrate it's balloon
> > > bitmap along with the migration data.
> > Nothing is special in the second migration, QEMU will request the
> > guest for free pages Information, and the guest will traverse it's
> > current free page list to construct a new free page bitmap and send it to
> QEMU. Just like in the first migration.
> > Liang
> > >
> > > Dave
> > >
> > > --
> > > Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK
> Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK