Linux Virtualization - Jun 2018 - [virtio-dev] Re: [PATCH v33 2/4] virtio-balloon: VIRTIO_BALLOON_F_FREE_PAGE_HINT

On Tue, Jun 19, 2018 at 01:06:48AM +0000, Wang, Wei W
wrote:
> On Monday, June 18, 2018 10:29 AM, Michael S. Tsirkin wrote:
> > On Sat, Jun 16, 2018 at 01:09:44AM +0000, Wang, Wei W wrote:
> > > Not necessarily, I think. We have min(4m_page_blocks / 512, 1024)
above,
> > so the maximum memory that can be reported is 2TB. For larger guests,
e.g.
> > 4TB, the optimization can still offer 2TB free memory (better than no
> > optimization).
> > 
> > Maybe it's better, maybe it isn't. It certainly muddies the
waters even more.
> > I'd rather we had a better plan. From that POV I like what Matthew
Wilcox
> > suggested for this which is to steal the necessary # of entries off
the list.
> 
> Actually what Matthew suggested doesn't make a difference here. That
method always steal the first free page blocks, and sure can be changed to take
more. But all these can be achieved via kmalloc
I'd do get_user_pages really. You don't want pages split, etc.
> by the caller which is more prudent and makes the code more
straightforward. I think we don't need to take that risk unless the MM folks
strongly endorse that approach.
> 
> The max size of the kmalloc-ed memory is 4MB, which gives us the limitation
that the max free memory to report is 2TB. Back to the motivation of this work,
the cloud guys want to use this optimization to accelerate their guest live
migration. 2TB guests are not common in today's clouds. When huge guests
become common in the future, we can easily tweak this API to fill hints into
scattered buffer (e.g. several 4MB arrays passed to this API) instead of one as
in this version.
> 
> This limitation doesn't cause any issue from functionality perspective.
For the extreme case like a 100TB guest live migration which is theoretically
possible today, this optimization helps skip 2TB of its free memory. This result
is that it may reduce only 2% live migration time, but still better than not
skipping the 2TB (if not using the feature).
Not clearly better, no, since you are slowing the guest.

> So, for the first release of this feature, I think it is better to have the
simpler and more straightforward solution as we have now, and clearly document
why it can report up to 2TB free memory.
No one has the time to read documentation about how an internal flag
within a device works. Come on, getting two pages isn't much harder
than a single one.
> 
>  
> > If that doesn't fly, we can allocate out of the loop and just
retry with more
> > pages.
> > 
> > > On the other hand, large guests being large mostly because the
guests need
> > to use large memory. In that case, they usually won't have that
much free
> > memory to report.
> > 
> > And following this logic small guests don't have a lot of memory
to report at
> > all.
> > Could you remind me why are we considering this optimization then?
> 
> If there is a 3TB guest, it is 3TB not 2TB mostly because it would need to
use e.g. 2.5TB memory from time to time. In the worst case, it only has 0.5TB
free memory to report, but reporting 0.5TB with this optimization is better than
no optimization. (and the current 2TB limitation isn't a limitation for the
3TB guest in this case)
I'd rather not spend time writing up random limitations.

> Best,
> Wei

On 06/19/2018 11:05 AM, Michael S. Tsirkin wrote:
> On Tue, Jun 19, 2018 at 01:06:48AM +0000, Wang, Wei W wrote:
>> On Monday, June 18, 2018 10:29 AM, Michael S. Tsirkin wrote:
>>> On Sat, Jun 16, 2018 at 01:09:44AM +0000, Wang, Wei W wrote:
>>>> Not necessarily, I think. We have min(4m_page_blocks / 512,
1024) above,
>>> so the maximum memory that can be reported is 2TB. For larger
guests, e.g.
>>> 4TB, the optimization can still offer 2TB free memory (better than
no
>>> optimization).
>>>
>>> Maybe it's better, maybe it isn't. It certainly muddies the
waters even more.
>>> I'd rather we had a better plan. From that POV I like what
Matthew Wilcox
>>> suggested for this which is to steal the necessary # of entries off
the list.
>> Actually what Matthew suggested doesn't make a difference here.
That method always steal the first free page blocks, and sure can be changed to
take more. But all these can be achieved via kmalloc
> I'd do get_user_pages really. You don't want pages split, etc.
>> by the caller which is more prudent and makes the code more
straightforward. I think we don't need to take that risk unless the MM folks
strongly endorse that approach.
>>
>> The max size of the kmalloc-ed memory is 4MB, which gives us the
limitation that the max free memory to report is 2TB. Back to the motivation of
this work, the cloud guys want to use this optimization to accelerate their
guest live migration. 2TB guests are not common in today's clouds. When huge
guests become common in the future, we can easily tweak this API to fill hints
into scattered buffer (e.g. several 4MB arrays passed to this API) instead of
one as in this version.
>>
>> This limitation doesn't cause any issue from functionality
perspective. For the extreme case like a 100TB guest live migration which is
theoretically possible today, this optimization helps skip 2TB of its free
memory. This result is that it may reduce only 2% live migration time, but still
better than not skipping the 2TB (if not using the feature).
> Not clearly better, no, since you are slowing the guest.
Not really. Live migration slows down the guest itself. It seems that 
the guest spends a little extra time reporting free pages, but in return 
the live migration time gets reduced a lot, which makes the guest endure 
less from live migration. (there is no drop of the workload performance 
when using the optimization in the tests)


>
>
>> So, for the first release of this feature, I think it is better to have
the simpler and more straightforward solution as we have now, and clearly
document why it can report up to 2TB free memory.
> No one has the time to read documentation about how an internal flag
> within a device works. Come on, getting two pages isn't much harder
> than a single one.
>>   
>>> If that doesn't fly, we can allocate out of the loop and just
retry with more
>>> pages.
>>>
>>>> On the other hand, large guests being large mostly because the
guests need
>>> to use large memory. In that case, they usually won't have that
much free
>>> memory to report.
>>>
>>> And following this logic small guests don't have a lot of
memory to report at
>>> all.
>>> Could you remind me why are we considering this optimization then?
>> If there is a 3TB guest, it is 3TB not 2TB mostly because it would need
to use e.g. 2.5TB memory from time to time. In the worst case, it only has 0.5TB
free memory to report, but reporting 0.5TB with this optimization is better than
no optimization. (and the current 2TB limitation isn't a limitation for the
3TB guest in this case)
> I'd rather not spend time writing up random limitations.
This is not a random limitation. It would be more clear to see the code. 
Also I'm not sure how get_user_pages could be used in our case, and what 
you meant by "getting two pages". I'll post out a new version, and
we
can discuss on the code.


Best,
Wei

On Tue, Jun 19, 2018 at 08:13:37PM +0800, Wei Wang
wrote:
> On 06/19/2018 11:05 AM, Michael S. Tsirkin wrote:
> > On Tue, Jun 19, 2018 at 01:06:48AM +0000, Wang, Wei W wrote:
> > > On Monday, June 18, 2018 10:29 AM, Michael S. Tsirkin wrote:
> > > > On Sat, Jun 16, 2018 at 01:09:44AM +0000, Wang, Wei W wrote:
> > > > > Not necessarily, I think. We have min(4m_page_blocks /
512, 1024) above,
> > > > so the maximum memory that can be reported is 2TB. For
larger guests, e.g.
> > > > 4TB, the optimization can still offer 2TB free memory
(better than no
> > > > optimization).
> > > > 
> > > > Maybe it's better, maybe it isn't. It certainly
muddies the waters even more.
> > > > I'd rather we had a better plan. From that POV I like
what Matthew Wilcox
> > > > suggested for this which is to steal the necessary # of
entries off the list.
> > > Actually what Matthew suggested doesn't make a difference
here. That method always steal the first free page blocks, and sure can be
changed to take more. But all these can be achieved via kmalloc
> > I'd do get_user_pages really. You don't want pages split, etc.
Oops sorry. I meant get_free_pages .
> 
> > > by the caller which is more prudent and makes the code more
straightforward. I think we don't need to take that risk unless the MM folks
strongly endorse that approach.
> > > 
> > > The max size of the kmalloc-ed memory is 4MB, which gives us the
limitation that the max free memory to report is 2TB. Back to the motivation of
this work, the cloud guys want to use this optimization to accelerate their
guest live migration. 2TB guests are not common in today's clouds. When huge
guests become common in the future, we can easily tweak this API to fill hints
into scattered buffer (e.g. several 4MB arrays passed to this API) instead of
one as in this version.
> > > 
> > > This limitation doesn't cause any issue from functionality
perspective. For the extreme case like a 100TB guest live migration which is
theoretically possible today, this optimization helps skip 2TB of its free
memory. This result is that it may reduce only 2% live migration time, but still
better than not skipping the 2TB (if not using the feature).
> > Not clearly better, no, since you are slowing the guest.
> 
> Not really. Live migration slows down the guest itself. It seems that the
> guest spends a little extra time reporting free pages, but in return the
> live migration time gets reduced a lot, which makes the guest endure less
> from live migration. (there is no drop of the workload performance when
> using the optimization in the tests)
My point was you can't say what is better without measuring.
Without special limitations you have hint overhead vs migration
overhead. I think we need to  build to scale to huge guests.
We might discover scalability problems down the road,
but no sense in building in limitations straight away.
> 
> 
> > 
> > 
> > > So, for the first release of this feature, I think it is better
to have the simpler and more straightforward solution as we have now, and
clearly document why it can report up to 2TB free memory.
> > No one has the time to read documentation about how an internal flag
> > within a device works. Come on, getting two pages isn't much
harder
> > than a single one.
> 
> > > > If that doesn't fly, we can allocate out of the loop and
just retry with more
> > > > pages.
> > > > 
> > > > > On the other hand, large guests being large mostly
because the guests need
> > > > to use large memory. In that case, they usually won't
have that much free
> > > > memory to report.
> > > > 
> > > > And following this logic small guests don't have a lot
of memory to report at
> > > > all.
> > > > Could you remind me why are we considering this optimization
then?
> > > If there is a 3TB guest, it is 3TB not 2TB mostly because it
would need to use e.g. 2.5TB memory from time to time. In the worst case, it
only has 0.5TB free memory to report, but reporting 0.5TB with this optimization
is better than no optimization. (and the current 2TB limitation isn't a
limitation for the 3TB guest in this case)
> > I'd rather not spend time writing up random limitations.
> 
> This is not a random limitation. It would be more clear to see the code.
Users don't see code though, that's the point.

Exporting internal limitations from code to users isn't great.

> Also I'm not sure how get_user_pages could be used in our case, and
what you
> meant by "getting two pages". I'll post out a new version,
and we can
> discuss on the code.
Sorry, I meant get_free_pages.
> 
> Best,
> Wei