lkml.org 
[lkml]   [2020]   [Mar]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v1 3/3] virtio-balloon: Switch back to OOM handler for VIRTIO_BALLOON_F_DEFLATE_ON_OOM
Tested-by: Tyler Sanderson <tysand@google.com>

Test setup: VM with 16 CPU, 64GB RAM. Running Debian 10. We have a 42
GB file full of random bytes that we continually cat to /dev/null.
This fills the page cache as the file is read. Meanwhile we trigger
the balloon to inflate, with a target size of 53 GB. This setup causes
the balloon inflation to pressure the page cache as the page cache is
also trying to grow. Afterwards we shrink the balloon back to zero (so
total deflate = total inflate).

Without patch (kernel 4.19.0-5):
Inflation never reaches the target until we stop the "cat file >
/dev/null" process. Total inflation time was 542 seconds. The longest
period that made no net forward progress was 315 seconds (see attached
graph).
Result of "grep balloon /proc/vmstat" after the test:
balloon_inflate 154828377
balloon_deflate 154828377

With patch (kernel 5.6.0-rc4+):
Total inflation duration was 63 seconds. No deflate-queue activity
occurs when pressuring the page-cache.
Result of "grep balloon /proc/vmstat" after the test:
balloon_inflate 12968539
balloon_deflate 12968539

Conclusion: This patch fixes the issue. In the test it reduced
inflate/deflate activity by 12x, and reduced inflation time by 8.6x.
But more importantly, if we hadn't killed the "grep balloon
/proc/vmstat" process then, without the patch, the inflation process
would never reach the target.

Attached is a png of a graph showing the problematic behavior without
this patch. It shows deflate-queue activity increasing linearly while
balloon size stays constant over the course of more than 8 minutes of
the test.


On Thu, Feb 20, 2020 at 7:29 PM Tyler Sanderson <tysand@google.com> wrote:
>
> Testing this patch is on my short-term TODO list, but I wasn't able to get to it this week. It is prioritized.
>
> In the meantime, I can anecdotally vouch that kernels before 4.19, the ones using the OOM notifier callback, have roughly 10x faster balloon inflation when pressuring the cache. So I anticipate this patch will return to that state and help my use case.
>
> I will try to post official measurements of this patch next week.
>
> On Sun, Feb 16, 2020 at 1:47 AM Michael S. Tsirkin <mst@redhat.com> wrote:
>>
>> On Fri, Feb 14, 2020 at 10:51:43AM +0100, David Hildenbrand wrote:
>> > On 05.02.20 17:34, David Hildenbrand wrote:
>> > > Commit 71994620bb25 ("virtio_balloon: replace oom notifier with shrinker")
>> > > changed the behavior when deflation happens automatically. Instead of
>> > > deflating when called by the OOM handler, the shrinker is used.
>> > >
>> > > However, the balloon is not simply some slab cache that should be
>> > > shrunk when under memory pressure. The shrinker does not have a concept of
>> > > priorities, so this behavior cannot be configured.
>> > >
>> > > There was a report that this results in undesired side effects when
>> > > inflating the balloon to shrink the page cache. [1]
>> > > "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."
>> > >
>> > > The name "deflate on OOM" makes it pretty clear when deflation should
>> > > happen - after other approaches to reclaim memory failed, not while
>> > > reclaiming. This allows to minimize the footprint of a guest - memory
>> > > will only be taken out of the balloon when really needed.
>> > >
>> > > Especially, a drop_slab() will result in the whole balloon getting
>> > > deflated - undesired. While handling it via the OOM handler might not be
>> > > perfect, it keeps existing behavior. If we want a different behavior, then
>> > > we need a new feature bit and document it properly (although, there should
>> > > be a clear use case and the intended effects should be well described).
>> > >
>> > > Keep using the shrinker for VIRTIO_BALLOON_F_FREE_PAGE_HINT, because
>> > > this has no such side effects. Always register the shrinker with
>> > > VIRTIO_BALLOON_F_FREE_PAGE_HINT now. We are always allowed to reuse free
>> > > pages that are still to be processed by the guest. The hypervisor takes
>> > > care of identifying and resolving possible races between processing a
>> > > hinting request and the guest reusing a page.
>> > >
>> > > In contrast to pre commit 71994620bb25 ("virtio_balloon: replace oom
>> > > notifier with shrinker"), don't add a moodule parameter to configure the
>> > > number of pages to deflate on OOM. Can be re-added if really needed.
>> > > Also, pay attention that leak_balloon() returns the number of 4k pages -
>> > > convert it properly in virtio_balloon_oom_notify().
>> > >
>> > > Note1: using the OOM handler is frowned upon, but it really is what we
>> > > need for this feature.
>> > >
>> > > Note2: without VIRTIO_BALLOON_F_MUST_TELL_HOST (iow, always with QEMU) we
>> > > could actually skip sending deflation requests to our hypervisor,
>> > > making the OOM path *very* simple. Besically freeing pages and
>> > > updating the balloon. If the communication with the host ever
>> > > becomes a problem on this call path.
>> > >
>> >
>> > @Michael, how to proceed with this?
>> >
>>
>> I'd like to see some reports that this helps people.
>> e.g. a tested-by tag.
>>
>> > --
>> > Thanks,
>> >
>> > David / dhildenb
>>
[unhandled content-type:image/png]
\
 
 \ /
  Last update: 2020-03-08 05:48    [W:0.121 / U:1.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site