Re: [RFC PATCH] mm, oom: introduce vm.sacrifice_hugepage_on_oom

On Wed 17-02-21 10:42:24, Eiichi Tsukata wrote:

Hi All,

Firstly, thank you for your careful review and attention to my patch
(and apologies for top-posting!).  Let me first explain why our use
case requires hugetlb over THP and then elaborate on the difficulty we
have to maintain the correct number of hugepages in the pool, finally
concluding with why the proposed approach would help us. Hopefully you
can extend it to other use cases and justify the proposal.

We use Linux to operate a KVM-based hypervisor. Using hugepages to
back VM memory significantly increases performance and density. Each
VM incurs a 4k regular page overhead which can vary drastically even
at runtime (eg. depending on network traffic). In addition, the
software doesn't know upfront if users will power on one large VM or
several small VMs.

To manage the varying balance of 4k pages vs. hugepages, we originally
leveraged THP. However, constant fragmentation due to VM power cycles,
the varying overhead I mentioned above, and other operations like
reconfiguration of NIC RX buffers resulted in two problems:
1) There were no guarantees hugepages would be used; and
2) Constant memory compaction incurred a measurable overhead.

Having a userspace service managing hugetlb gave us significant
performance advantages and much needed determinism. It chooses when to
try and create more hugepages as well as how many hugepages to go
after. Elements like how many hugepages it actually gets, combined
with what operations are happening on the host, allow our service to
make educated decisions about when to compact memory, drop caches, and
retry growing (or shrinking) the pool.

OK, thanks for the clarification. Just to make sure I understand. This
means that you are pro-activelly and optimistically pre-allocate hugetlb
pages even when there is no immediate need for those, right?

But that comes with a challenge: despite listening on cgroup for
pressure notifications (which happen from those runtime events we do
not control),

We do also have global pressure (PSI) counters. Have you tried to look
into those and try to back off even when the situation becomes critical?

the service is not guaranteed to sacrifice hugepages
fast enough and that causes an OOM. The killer will normally take out
a VM even if there are plenty of unused hugepages and that's obviously
disruptive for users. For us, free hugepages are almost always expendable.

For the bloat cases which are predictable, a memory management service
can adjust the hugepage pool size ahead of time. But it can be hard to
anticipate all scenarios, and some can be very volatile. Having a
failsafe mechanism as proposed in this patch offers invaluable
protection when things are missed.

The proposal solves this problem by sacrificing hugepages inline even
when the pressure comes from kernel allocations. The userspace service
can later readjust the pool size without being under pressure. Given
this is configurable, and defaults to being off, we thought it would
be a nice addition to the kernel and appreciated by other users that
may have similar requirements.

Thanks for your usecase description. It helped me to understand what you
are doing and how this can be really useful for your particular setup.
This is really a very specific situation from my POV. I am not yet sure
this is generic enough to warrant for a yet another tunable. One thing
you can do [1] is to
hook into oom notifiers interface (register_oom_notifier) and release
pages from the callback. Why is that batter than a global tunable?
For one thing you can make the implementation tailored to your specific
usecase. As the review feedback has shown this would be more tricky to
be done in a general case. Unlike a generic solution it would allow you
to coordinate with your userspace if you need. Would something like that
work for you?

---
[1] and I have to say I hate myself for suggesting that because I was
really hoping this interface would go away. But the reality disagrees so
I gave up on that goal...
-- 
Michal Hocko
SUSE Labs