Re: Balloon pressuring page cache

[David Hildenbrand <david@redhat.com>]

On 05.02.20 08:35, Nadav Amit wrote:
>> On Feb 3, 2020, at 2:50 PM, Nadav Amit <namit@vmware.com> wrote:
>>
>>> On Feb 3, 2020, at 8:34 AM, David Hildenbrand <david@redhat.com> wrote:
>>>
>>> On 03.02.20 17:18, Alexander Duyck wrote:
>>>> On Mon, 2020-02-03 at 08:11 -0500, Michael S. Tsirkin wrote:
>>>>> On Thu, Jan 30, 2020 at 11:59:46AM -0800, Tyler Sanderson wrote:
>>>>>> On Thu, Jan 30, 2020 at 7:31 AM Wang, Wei W <wei.w.wang@intel.com> wrote:
>>>>>>
>>>>>>   On Thursday, January 30, 2020 11:03 PM, David Hildenbrand wrote:
>>>>>>> On 29.01.20 20:11, Tyler Sanderson wrote:
>>>>>>>> On Wed, Jan 29, 2020 at 2:31 AM David Hildenbrand <david@redhat.com
>>>>>>>> <mailto:david@redhat.com>> wrote:
>>>>>>>>
>>>>>>>>   On 29.01.20 01:22, Tyler Sanderson via Virtualization wrote:
>>>>>>>>> A primary advantage of virtio balloon over other memory reclaim
>>>>>>>>> mechanisms is that it can pressure the guest's page cache into
>>>>>>>>   shrinking.
>>>>>>>>> However, since the balloon driver changed to using the shrinker
>>>>>>   API
>>>>>>> <https://github.com/torvalds/linux/commit/71994620bb25a8b109388fefa9
>>>>>>> e99a28e355255a#diff-fd202acf694d9eba19c8c64da3e480c9> this
>>>>>>>>> use case has become a bit more tricky. I'm wondering what the
>>>>>>> intended
>>>>>>>>> device implementation is.
>>>>>>>>>
>>>>>>>>> When inflating the balloon against page cache (i.e. no free
>>>>>>   memory
>>>>>>>>> remains) vmscan.c will both shrink page cache, but also invoke
>>>>>>   the
>>>>>>>>> shrinkers -- including the balloon's shrinker. So the balloon
>>>>>>   driver
>>>>>>>>> allocates memory which requires reclaim, vmscan gets this memory
>>>>>>> by
>>>>>>>>> shrinking the balloon, and then the driver adds the memory back
>>>>>>   to
>>>>>>> the
>>>>>>>>> balloon. Basically a busy no-op.
>>>>>>
>>>>>>   Per my understanding, the balloon allocation won’t invoke shrinker as
>>>>>>   __GFP_DIRECT_RECLAIM isn't set, no?
>>>>>>
>>>>>> I could be wrong about the mechanism, but the device sees lots of activity on
>>>>>> the deflate queue. The balloon is being shrunk. And this only starts once all
>>>>>> free memory is depleted and we're inflating into page cache.
>>>>>
>>>>> So given this looks like a regression, maybe we should revert the
>>>>> patch in question 71994620bb25 ("virtio_balloon: replace oom notifier with shrinker")
>>>>> Besides, with VIRTIO_BALLOON_F_FREE_PAGE_HINT
>>>>> shrinker also ignores VIRTIO_BALLOON_F_MUST_TELL_HOST which isn't nice
>>>>> at all.
>>>>>
>>>>> So it looks like all this rework introduced more issues than it
>>>>> addressed ...
>>>>>
>>>>> I also CC Alex Duyck for an opinion on this.
>>>>> Alex, what do you use to put pressure on page cache?
>>>>
>>>> I would say reverting probably makes sense. I'm not sure there is much
>>>> value to having a shrinker running deflation when you are actively trying
>>>> to increase the balloon. It would make more sense to wait until you are
>>>> actually about to start hitting oom.
>>>
>>> I think the shrinker makes sense for free page hinting feature
>>> (everything on free_page_list).
>>>
>>> So instead of only reverting, I think we should split it up and always
>>> register the shrinker for VIRTIO_BALLOON_F_FREE_PAGE_HINT and the OOM
>>> notifier (as before) for VIRTIO_BALLOON_F_MUST_TELL_HOST.
>>>
>>> (Of course, adapting what is being done in the shrinker and in the OOM
>>> notifier)
>>
>> David,
>>
>> Please keep me posted. I decided to adapt the same solution as the virtio
>> balloon for the VMware balloon. If the verdict is that this is damaging and
>> the OOM notifier should be used instead, I will submit patches to move to
>> OOM notifier as well.
> 
> Adding some information for the record (if someone googles this thread):
> 
> In the VMware balloon driver, the shrinker is disabled by default since we
> encountered a performance degradation in testing. I tried to avoid rapid
> inflation/shrinker-deflation cycles by adding a timeout, but apparently it
> did not help in avoiding the performance regression.

Thanks for that info. To me that sounds like the shrinker is the wrong
approach to "auto-deflation". It's not just "some slab cache".


-- 
Thanks,

David / dhildenb