Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1352892 > unrolled thread

Re: [RFC qemu 0/4] A PV solution for live migration optimization

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [RFC qemu 0/4] A PV solution for live migration optimization Amit Shah <amit.shah@redhat.com> - 2016-03-08 12:20 +0100
    RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-08 14:20 +0100
    RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-10 08:50 +0100
      Re: [RFC qemu 0/4] A PV solution for live migration optimization Amit Shah <amit.shah@redhat.com> - 2016-03-10 09:00 +0100
        RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-10 09:40 +0100
          Re: [RFC qemu 0/4] A PV solution for live migration optimization "Dr. David Alan Gilbert" <dgilbert@redhat.com> - 2016-03-10 12:20 +0100
            RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-11 03:40 +0100
              Re: [RFC qemu 0/4] A PV solution for live migration optimization "Dr. David Alan Gilbert" <dgilbert@redhat.com> - 2016-03-14 18:10 +0100
                RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-15 04:40 +0100
                Re: [RFC qemu 0/4] A PV solution for live migration optimization "Michael S. Tsirkin" <mst@redhat.com> - 2016-03-15 11:40 +0100
                  RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-15 12:20 +0100
                    Re: [RFC qemu 0/4] A PV solution for live migration optimization "Dr. David Alan Gilbert" <dgilbert@redhat.com> - 2016-03-15 21:00 +0100
                      RE: [RFC qemu 0/4] A PV solution for live migration optimization "Li, Liang Z" <liang.z.li@intel.com> - 2016-03-16 02:30 +0100

#1352892 — Re: [RFC qemu 0/4] A PV solution for live migration optimization

On (Thu) 03 Mar 2016 [18:44:24], Liang Li wrote:
> The current QEMU live migration implementation mark the all the
> guest's RAM pages as dirtied in the ram bulk stage, all these pages
> will be processed and that takes quit a lot of CPU cycles.
> 
> From guest's point of view, it doesn't care about the content in free
> pages. We can make use of this fact and skip processing the free
> pages in the ram bulk stage, it 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 QEMU side implementation.
> 
> The virtio-balloon is extended so that QEMU can get the free pages
> information from the guest through virtio.
> 
> After getting the free pages information (a bitmap), QEMU can use it
> to filter out the guest's free pages in the ram bulk stage. This make
> the live migration process much more efficient.
> 
> This RFC version doesn't take the post-copy and RDMA into
> consideration, maybe both of them can benefit from this PV solution
> by with some extra modifications.

I like the idea, just have to prove (review) and test it a lot to
ensure we don't end up skipping pages that matter.

However, there are a couple of points:

In my opinion, the information that's exchanged between the guest and
the host should be exchanged over a virtio-serial channel rather than
virtio-balloon.  First, there's nothing related to the balloon here.
It just happens to be memory info.  Second, I would never enable
balloon in a guest that I want to be performance-sensitive.  So even
if you add this as part of balloon, you'll find no one is using this
solution.

Secondly, I suggest virtio-serial, because it's meant exactly to
exchange free-flowing information between a host and a guest, and you
don't need to extend any part of the protocol for it (hence no changes
necessary to the spec).  You can see how spice, vnc, etc., use
virtio-serial to exchange data.


		Amit

[toc] | [next] | [standalone]

#1353033

> Subject: Re: [RFC qemu 0/4] A PV solution for live migration optimization
> 
> On (Thu) 03 Mar 2016 [18:44:24], Liang Li wrote:
> > The current QEMU live migration implementation mark the all the
> > guest's RAM pages as dirtied in the ram bulk stage, all these pages
> > will be processed and that takes quit a lot of CPU cycles.
> >
> > From guest's point of view, it doesn't care about the content in free
> > pages. We can make use of this fact and skip processing the free pages
> > in the ram bulk stage, it 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 QEMU side implementation.
> >
> > The virtio-balloon is extended so that QEMU can get the free pages
> > information from the guest through virtio.
> >
> > After getting the free pages information (a bitmap), QEMU can use it
> > to filter out the guest's free pages in the ram bulk stage. This make
> > the live migration process much more efficient.
> >
> > This RFC version doesn't take the post-copy and RDMA into
> > consideration, maybe both of them can benefit from this PV solution by
> > with some extra modifications.
> 
> I like the idea, just have to prove (review) and test it a lot to ensure we don't
> end up skipping pages that matter.
> 
> However, there are a couple of points:
> 
> In my opinion, the information that's exchanged between the guest and the
> host should be exchanged over a virtio-serial channel rather than virtio-
> balloon.  First, there's nothing related to the balloon here.
> It just happens to be memory info.  Second, I would never enable balloon in
> a guest that I want to be performance-sensitive.  So even if you add this as
> part of balloon, you'll find no one is using this solution.
> 
> Secondly, I suggest virtio-serial, because it's meant exactly to exchange free-
> flowing information between a host and a guest, and you don't need to
> extend any part of the protocol for it (hence no changes necessary to the
> spec).  You can see how spice, vnc, etc., use virtio-serial to exchange data.
> 
> 
> 		Amit

I don't like to use the virtio-balloon too, and it's confusing. 
It's grate if the virtio-serial can be used, I will take a look at it. 

Thanks for your suggestion!

Liang

[toc] | [prev] | [next] | [standalone]

#1354902

> > This patch set is the QEMU side implementation.
> >
> > The virtio-balloon is extended so that QEMU can get the free pages
> > information from the guest through virtio.
> >
> > After getting the free pages information (a bitmap), QEMU can use it
> > to filter out the guest's free pages in the ram bulk stage. This make
> > the live migration process much more efficient.
> >
> > This RFC version doesn't take the post-copy and RDMA into
> > consideration, maybe both of them can benefit from this PV solution by
> > with some extra modifications.
> 
> I like the idea, just have to prove (review) and test it a lot to ensure we don't
> end up skipping pages that matter.
> 
> However, there are a couple of points:
> 
> In my opinion, the information that's exchanged between the guest and the
> host should be exchanged over a virtio-serial channel rather than virtio-
> balloon.  First, there's nothing related to the balloon here.
> It just happens to be memory info.  Second, I would never enable balloon in
> a guest that I want to be performance-sensitive.  So even if you add this as
> part of balloon, you'll find no one is using this solution.
> 
> Secondly, I suggest virtio-serial, because it's meant exactly to exchange free-
> flowing information between a host and a guest, and you don't need to
> extend any part of the protocol for it (hence no changes necessary to the
> spec).  You can see how spice, vnc, etc., use virtio-serial to exchange data.
> 
> 
> 		Amit

Hi Amit,

 Could provide more information on how to use virtio-serial to exchange data?  Thread , Wiki or code are all OK. 
 I have not find some useful information yet.

Thanks
Liang

[toc] | [prev] | [next] | [standalone]

#1354906

On (Thu) 10 Mar 2016 [07:44:19], Li, Liang Z wrote:
> 
> Hi Amit,
> 
>  Could provide more information on how to use virtio-serial to exchange data?  Thread , Wiki or code are all OK. 
>  I have not find some useful information yet.

See this commit in the Linux sources:

108fc82596e3b66b819df9d28c1ebbc9ab5de14c

that adds a way to send guest trace data over to the host.  I think
that's the most relevant to your use-case.  However, you'll have to
add an in-kernel user of virtio-serial (like the virtio-console code
-- the code that deals with tty and hvc currently).  There's no other
non-tty user right now, and this is the right kind of use-case to add
one for!

For many other (userspace) use-cases, see the qemu-guest-agent in the
qemu sources.

The API is documented in the wiki:

http://www.linux-kvm.org/page/Virtio-serial_API

and the feature pages have some information that may help as well:

https://fedoraproject.org/wiki/Features/VirtioSerial

There are some links in here too:

http://log.amitshah.net/2010/09/communication-between-guests-and-hosts/

Hope this helps.


		Amit

[toc] | [prev] | [next] | [standalone]

#1354920

> >  Could provide more information on how to use virtio-serial to exchange
> data?  Thread , Wiki or code are all OK.
> >  I have not find some useful information yet.
> 
> See this commit in the Linux sources:
> 
> 108fc82596e3b66b819df9d28c1ebbc9ab5de14c
> 
> that adds a way to send guest trace data over to the host.  I think that's the
> most relevant to your use-case.  However, you'll have to add an in-kernel
> user of virtio-serial (like the virtio-console code
> -- the code that deals with tty and hvc currently).  There's no other non-tty
> user right now, and this is the right kind of use-case to add one for!
> 
> For many other (userspace) use-cases, see the qemu-guest-agent in the
> qemu sources.
> 
> The API is documented in the wiki:
> 
> http://www.linux-kvm.org/page/Virtio-serial_API
> 
> and the feature pages have some information that may help as well:
> 
> https://fedoraproject.org/wiki/Features/VirtioSerial
> 
> There are some links in here too:
> 
> http://log.amitshah.net/2010/09/communication-between-guests-and-
> hosts/
> 
> Hope this helps.
> 
> 
> 		Amit

Thanks a lot !!

Liang

[toc] | [prev] | [next] | [standalone]

#1355038

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.
     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?

  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 migration;
     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 your
     guest/network/reqirements.

  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?

       * 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?

  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?

  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.
     
Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1355553

> 
> 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, right?

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.

>   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 migration;
>      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 your
>      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.

>        * 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.

>   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?

>   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@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1357426

* Li, Liang Z (liang.z.li@intel.com) wrote:
> > 
> > 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, right?
> 
> 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 migration;
> >      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 your
> >      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.

> >        * 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.

> >   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.

Right.

Dave

> Liang
> > 
> > Dave
> > 
> > --
> > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1357791

> > > 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,
> right?
> >
> > 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
> migration;
> > >      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
> your
> > >      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
by guest.

> > >   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.
> 
> Right.
> 
> Dave
> 
> > Liang
> > >
> > > Dave
> > >
> > > --
> > > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
> --
> Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1357976

On Mon, Mar 14, 2016 at 05:03:34PM +0000, Dr. David Alan Gilbert wrote:
> * Li, Liang Z (liang.z.li@intel.com) wrote:
> > > 
> > > 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, right?
> > 
> > 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.

This works once, sure. But there's an issue is that you have
to defer migration until you get the free page list,
and this only works once. So you end up with heuristics
about how long to wait.

Instead I propose:

- mark all pages dirty as we do now.

- at start of migration, start tracking dirty
  pages in kvm, and tell guest to start tracking free pages

we can now introduce any kind of delay, for
example wait for ack from guest, or do whatever else,
or even just start migrating pages

- repeatedly:
	- get list of free pages from guest
	- clear them in migration bitmap
	- get dirty list from kvm

- at end of migration, stop tracking writes in kvm,
  and tell guest to stop tracking free pages



> > >   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 migration;
> > >      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 your
> > >      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.
> 
> > >        * 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.
> 
> > >   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.
> 
> Right.
> 
> Dave
> 
> > Liang
> > > 
> > > Dave
> > > 
> > > --
> > > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
> --
> Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1357993

> On Mon, Mar 14, 2016 at 05:03:34PM +0000, Dr. David Alan Gilbert wrote:
> > * Li, Liang Z (liang.z.li@intel.com) wrote:
> > > >
> > > > 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,
> right?
> > >
> > > 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.
> 
> This works once, sure. But there's an issue is that you have to defer migration
> until you get the free page list, and this only works once. So you end up with
> heuristics about how long to wait.
> 
> Instead I propose:
> 
> - mark all pages dirty as we do now.
> 
> - at start of migration, start tracking dirty
>   pages in kvm, and tell guest to start tracking free pages
> 
> we can now introduce any kind of delay, for example wait for ack from guest,
> or do whatever else, or even just start migrating pages
> 
> - repeatedly:
> 	- get list of free pages from guest
> 	- clear them in migration bitmap
> 	- get dirty list from kvm
> 
> - at end of migration, stop tracking writes in kvm,
>   and tell guest to stop tracking free pages

I had thought of filtering out the free pages in each migration bitmap synchronization. 
The advantage is we can skip process as many free pages as possible. Not just once.
The disadvantage is that we should change the current memory management code to track the free pages,
instead of traversing the free page list to construct the free pages bitmap, to reduce the overhead to get the free pages bitmap.
I am not sure the if the Kernel people would like it.

If keeping the traversing mechanism, because of the overhead, maybe it's not worth to filter out the free pages repeatedly.

Liang

[toc] | [prev] | [next] | [standalone]

#1358220

* Li, Liang Z (liang.z.li@intel.com) wrote:
> > On Mon, Mar 14, 2016 at 05:03:34PM +0000, Dr. David Alan Gilbert wrote:
> > > * Li, Liang Z (liang.z.li@intel.com) wrote:
> > > > >
> > > > > 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,
> > right?
> > > >
> > > > 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.
> > 
> > This works once, sure. But there's an issue is that you have to defer migration
> > until you get the free page list, and this only works once. So you end up with
> > heuristics about how long to wait.
> > 
> > Instead I propose:
> > 
> > - mark all pages dirty as we do now.
> > 
> > - at start of migration, start tracking dirty
> >   pages in kvm, and tell guest to start tracking free pages
> > 
> > we can now introduce any kind of delay, for example wait for ack from guest,
> > or do whatever else, or even just start migrating pages
> > 
> > - repeatedly:
> > 	- get list of free pages from guest
> > 	- clear them in migration bitmap
> > 	- get dirty list from kvm
> > 
> > - at end of migration, stop tracking writes in kvm,
> >   and tell guest to stop tracking free pages
> 
> I had thought of filtering out the free pages in each migration bitmap synchronization. 
> The advantage is we can skip process as many free pages as possible. Not just once.
> The disadvantage is that we should change the current memory management code to track the free pages,
> instead of traversing the free page list to construct the free pages bitmap, to reduce the overhead to get the free pages bitmap.
> I am not sure the if the Kernel people would like it.
> 
> If keeping the traversing mechanism, because of the overhead, maybe it's not worth to filter out the free pages repeatedly.

Well, Michael's idea of not waiting for the dirty
bitmap to be filled does make that idea of constnatly
using the free-bitmap better.

In that case, is it easier if something (guest/host?)
allocates some memory in the guests physical RAM space
and just points the host to it, rather than having an 
explicit 'send'.

Dave

> Liang
> 
> 
> 
> 
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

[toc] | [prev] | [next] | [standalone]

#1358464

> > > > > >   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,
> > > right?
> > > > >
> > > > > 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.
> > >
> > > This works once, sure. But there's an issue is that you have to
> > > defer migration until you get the free page list, and this only
> > > works once. So you end up with heuristics about how long to wait.
> > >
> > > Instead I propose:
> > >
> > > - mark all pages dirty as we do now.
> > >
> > > - at start of migration, start tracking dirty
> > >   pages in kvm, and tell guest to start tracking free pages
> > >
> > > we can now introduce any kind of delay, for example wait for ack
> > > from guest, or do whatever else, or even just start migrating pages
> > >
> > > - repeatedly:
> > > 	- get list of free pages from guest
> > > 	- clear them in migration bitmap
> > > 	- get dirty list from kvm
> > >
> > > - at end of migration, stop tracking writes in kvm,
> > >   and tell guest to stop tracking free pages
> >
> > I had thought of filtering out the free pages in each migration bitmap
> synchronization.
> > The advantage is we can skip process as many free pages as possible. Not
> just once.
> > The disadvantage is that we should change the current memory
> > management code to track the free pages, instead of traversing the free
> page list to construct the free pages bitmap, to reduce the overhead to get
> the free pages bitmap.
> > I am not sure the if the Kernel people would like it.
> >
> > If keeping the traversing mechanism, because of the overhead, maybe it's
> not worth to filter out the free pages repeatedly.
> 
> Well, Michael's idea of not waiting for the dirty bitmap to be filled does make
> that idea of constnatly using the free-bitmap better.
> 

No wait is a good idea.
Actually, we could shorten the waiting time by pre allocating the free pages bit map
and update it when guest allocating/freeing pages. it requires to modify the mm 
related code. I don't know whether the kernel people like this.

> In that case, is it easier if something (guest/host?) allocates some memory in
> the guests physical RAM space and just points the host to it, rather than
> having an explicit 'send'.
> 

Good idea too.

Liang
> Dave

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web