Virtualization - Aug 2018 - [PATCH v2 2/2] virtio_balloon: replace oom notifier with shrinker

On Wed 01-08-18 19:12:25, Wei Wang wrote:
> On 07/30/2018 05:00 PM, Michal Hocko wrote:
> > On Fri 27-07-18 17:24:55, Wei Wang wrote:
> > > The OOM notifier is getting deprecated to use for the reasons
mentioned
> > > here by Michal Hocko: https://lkml.org/lkml/2018/7/12/314
> > > 
> > > This patch replaces the virtio-balloon oom notifier with a
shrinker
> > > to release balloon pages on memory pressure.
> > It would be great to document the replacement. This is not a small
> > change...
> 
> OK. I plan to document the following to the commit log:
> 
>   The OOM notifier is getting deprecated to use for the reasons:
>     - As a callout from the oom context, it is too subtle and easy to
>       generate bugs and corner cases which are hard to track;
>     - It is called too late (after the reclaiming has been performed).
>       Drivers with large amuont of reclaimable memory is expected to be
>       released them at an early age of memory pressure;
>     - The notifier callback isn't aware of the oom contrains;
>     Link: https://lkml.org/lkml/2018/7/12/314
> 
>     This patch replaces the virtio-balloon oom notifier with a shrinker
>     to release balloon pages on memory pressure. Users can set the amount
of
>     memory pages to release each time a shrinker_scan is called via the
>     module parameter balloon_pages_to_shrink, and the default amount is 256
>     pages. Historically, the feature VIRTIO_BALLOON_F_DEFLATE_ON_OOM has
>     been used to release balloon pages on OOM. We continue to use this
>     feature bit for the shrinker, so the shrinker is only registered when
>     this feature bit has been negotiated with host.
Do you have any numbers for how does this work in practice? Let's say
you have a medium page cache workload which triggers kswapd to do a
light reclaim? Hardcoded shrinking sounds quite dubious to me but I have
no idea how people expect this to work. Shouldn't this be more
adaptive? How precious are those pages anyway?
-- 
Michal Hocko
SUSE Labs

On 08/01/2018 07:34 PM, Michal Hocko wrote:
> On Wed 01-08-18 19:12:25, Wei Wang wrote:
>> On 07/30/2018 05:00 PM, Michal Hocko wrote:
>>> On Fri 27-07-18 17:24:55, Wei Wang wrote:
>>>> The OOM notifier is getting deprecated to use for the reasons
mentioned
>>>> here by Michal Hocko: https://lkml.org/lkml/2018/7/12/314
>>>>
>>>> This patch replaces the virtio-balloon oom notifier with a
shrinker
>>>> to release balloon pages on memory pressure.
>>> It would be great to document the replacement. This is not a small
>>> change...
>> OK. I plan to document the following to the commit log:
>>
>>    The OOM notifier is getting deprecated to use for the reasons:
>>      - As a callout from the oom context, it is too subtle and easy to
>>        generate bugs and corner cases which are hard to track;
>>      - It is called too late (after the reclaiming has been performed).
>>        Drivers with large amuont of reclaimable memory is expected to
be
>>        released them at an early age of memory pressure;
>>      - The notifier callback isn't aware of the oom contrains;
>>      Link: https://lkml.org/lkml/2018/7/12/314
>>
>>      This patch replaces the virtio-balloon oom notifier with a
shrinker
>>      to release balloon pages on memory pressure. Users can set the
amount of
>>      memory pages to release each time a shrinker_scan is called via
the
>>      module parameter balloon_pages_to_shrink, and the default amount
is 256
>>      pages. Historically, the feature VIRTIO_BALLOON_F_DEFLATE_ON_OOM
has
>>      been used to release balloon pages on OOM. We continue to use this
>>      feature bit for the shrinker, so the shrinker is only registered
when
>>      this feature bit has been negotiated with host.
> Do you have any numbers for how does this work in practice?
It works in this way: for example, we can set the parameter, 
balloon_pages_to_shrink, to shrink 1GB memory once shrink scan is 
called. Now, we have a 8GB guest, and we balloon out 7GB. When shrink 
scan is called, the balloon driver will get back 1GB memory and give 
them back to mm, then the ballooned memory becomes 6GB.

When the shrinker scan is called the second time, another 1GB will be 
given back to mm. So the ballooned pages are given back to mm gradually.
> Let's say
> you have a medium page cache workload which triggers kswapd to do a
> light reclaim? Hardcoded shrinking sounds quite dubious to me but I have
> no idea how people expect this to work. Shouldn't this be more
> adaptive? How precious are those pages anyway?
Those pages are given to host to use usually because the guest has 
enough free memory, and host doesn't want to waste those pieces of 
memory as they are not used by this guest. When the guest needs them, it 
is reasonable that the guest has higher priority to take them back.
But I'm not sure if there would be a more adaptive approach than 
"gradually giving back as the guest wants more".

Best,
Wei

On 2018/08/02 19:32, Wei Wang wrote:
> On 08/01/2018 07:34 PM, Michal Hocko wrote:
>> Do you have any numbers for how does this work in practice?
> 
> It works in this way: for example, we can set the parameter,
balloon_pages_to_shrink,
> to shrink 1GB memory once shrink scan is called. Now, we have a 8GB guest,
and we balloon
> out 7GB. When shrink scan is called, the balloon driver will get back 1GB
memory and give
> them back to mm, then the ballooned memory becomes 6GB.
Since shrinker might be called concurrently (am I correct?), the balloon might
deflate
far more than needed if it releases such much memory. If shrinker is used,
releasing 256
pages might be sufficient.

On Thu 02-08-18 18:32:44, Wei Wang wrote:
> On 08/01/2018 07:34 PM, Michal Hocko wrote:
> > On Wed 01-08-18 19:12:25, Wei Wang wrote:
> > > On 07/30/2018 05:00 PM, Michal Hocko wrote:
> > > > On Fri 27-07-18 17:24:55, Wei Wang wrote:
> > > > > The OOM notifier is getting deprecated to use for the
reasons mentioned
> > > > > here by Michal Hocko:
https://lkml.org/lkml/2018/7/12/314
> > > > > 
> > > > > This patch replaces the virtio-balloon oom notifier
with a shrinker
> > > > > to release balloon pages on memory pressure.
> > > > It would be great to document the replacement. This is not a
small
> > > > change...
> > > OK. I plan to document the following to the commit log:
> > > 
> > >    The OOM notifier is getting deprecated to use for the reasons:
> > >      - As a callout from the oom context, it is too subtle and
easy to
> > >        generate bugs and corner cases which are hard to track;
> > >      - It is called too late (after the reclaiming has been
performed).
> > >        Drivers with large amuont of reclaimable memory is
expected to be
> > >        released them at an early age of memory pressure;
> > >      - The notifier callback isn't aware of the oom
contrains;
> > >      Link: https://lkml.org/lkml/2018/7/12/314
> > > 
> > >      This patch replaces the virtio-balloon oom notifier with a
shrinker
> > >      to release balloon pages on memory pressure. Users can set
the amount of
> > >      memory pages to release each time a shrinker_scan is called
via the
> > >      module parameter balloon_pages_to_shrink, and the default
amount is 256
> > >      pages. Historically, the feature
VIRTIO_BALLOON_F_DEFLATE_ON_OOM has
> > >      been used to release balloon pages on OOM. We continue to
use this
> > >      feature bit for the shrinker, so the shrinker is only
registered when
> > >      this feature bit has been negotiated with host.
> > Do you have any numbers for how does this work in practice?
> 
> It works in this way: for example, we can set the parameter,
> balloon_pages_to_shrink, to shrink 1GB memory once shrink scan is called.
> Now, we have a 8GB guest, and we balloon out 7GB. When shrink scan is
> called, the balloon driver will get back 1GB memory and give them back to
> mm, then the ballooned memory becomes 6GB.
> 
> When the shrinker scan is called the second time, another 1GB will be given
> back to mm. So the ballooned pages are given back to mm gradually.
> 
> > Let's say
> > you have a medium page cache workload which triggers kswapd to do a
> > light reclaim? Hardcoded shrinking sounds quite dubious to me but I
have
> > no idea how people expect this to work. Shouldn't this be more
> > adaptive? How precious are those pages anyway?
> 
> Those pages are given to host to use usually because the guest has enough
> free memory, and host doesn't want to waste those pieces of memory as
they
> are not used by this guest. When the guest needs them, it is reasonable
that
> the guest has higher priority to take them back.
> But I'm not sure if there would be a more adaptive approach than
"gradually
> giving back as the guest wants more".
I am not sure I follow. Let me be more specific. Say you have a trivial
stream IO triggering reclaim to recycle clean page cache. This will
invoke slab shrinkers as well. Do you really want to drop your batch of
pages on each invocation? Doesn't that remove them very quickly? Just
try to dd if=large_file of=/dev/null and see how your pages are
disappearing. Shrinkers usually scale the number of objects they are
going to reclaim based on the memory pressure (aka targer to be
reclaimed).
-- 
Michal Hocko
SUSE Labs

On Thu, Aug 02, 2018 at 06:32:44PM +0800, Wei Wang
wrote:
> On 08/01/2018 07:34 PM, Michal Hocko wrote:
> > On Wed 01-08-18 19:12:25, Wei Wang wrote:
> > > On 07/30/2018 05:00 PM, Michal Hocko wrote:
> > > > On Fri 27-07-18 17:24:55, Wei Wang wrote:
> > > > > The OOM notifier is getting deprecated to use for the
reasons mentioned
> > > > > here by Michal Hocko:
https://lkml.org/lkml/2018/7/12/314
> > > > > 
> > > > > This patch replaces the virtio-balloon oom notifier
with a shrinker
> > > > > to release balloon pages on memory pressure.
> > > > It would be great to document the replacement. This is not a
small
> > > > change...
> > > OK. I plan to document the following to the commit log:
> > > 
> > >    The OOM notifier is getting deprecated to use for the reasons:
> > >      - As a callout from the oom context, it is too subtle and
easy to
> > >        generate bugs and corner cases which are hard to track;
> > >      - It is called too late (after the reclaiming has been
performed).
> > >        Drivers with large amuont of reclaimable memory is
expected to be
> > >        released them at an early age of memory pressure;
> > >      - The notifier callback isn't aware of the oom
contrains;
> > >      Link: https://lkml.org/lkml/2018/7/12/314
> > > 
> > >      This patch replaces the virtio-balloon oom notifier with a
shrinker
> > >      to release balloon pages on memory pressure. Users can set
the amount of
> > >      memory pages to release each time a shrinker_scan is called
via the
> > >      module parameter balloon_pages_to_shrink, and the default
amount is 256
> > >      pages. Historically, the feature
VIRTIO_BALLOON_F_DEFLATE_ON_OOM has
> > >      been used to release balloon pages on OOM. We continue to
use this
> > >      feature bit for the shrinker, so the shrinker is only
registered when
> > >      this feature bit has been negotiated with host.
> > Do you have any numbers for how does this work in practice?
> 
> It works in this way: for example, we can set the parameter,
> balloon_pages_to_shrink, to shrink 1GB memory once shrink scan is called.
> Now, we have a 8GB guest, and we balloon out 7GB. When shrink scan is
> called, the balloon driver will get back 1GB memory and give them back to
> mm, then the ballooned memory becomes 6GB.
> 
> When the shrinker scan is called the second time, another 1GB will be given
> back to mm. So the ballooned pages are given back to mm gradually.
I think what's being asked here is a description of tests that
were run. Which workloads see improved behaviour?

Our behaviour under memory pressure isn't great, in particular it is not
clear when it's safe to re-inflate the balloon, if host attempts to
re-inflate it too soon then we still get OOM. It would be better
if VIRTIO_BALLOON_F_DEFLATE_ON_OOM would somehow mean
"it's ok to ask for almost all of memory, if guest needs memory from
balloon for apps to function it can take it from the balloon".


-- 
MST