Re: [RFC] Mechanism to induce memory reclaim

[Johannes Weiner <hannes@cmpxchg.org>]

On Mon, Mar 07, 2022 at 06:31:41PM +0000, Shakeel Butt wrote:
> On Mon, Mar 07, 2022 at 03:41:45PM +0100, Michal Hocko wrote:
> > On Sun 06-03-22 15:11:23, David Rientjes wrote:
> > [...]
> > > Some questions to get discussion going:
> > >
> > >  - Overall feedback or suggestions for the proposal in general?
> 
> > Do we really need this interface? What would be usecases which cannot
> > use an existing interfaces we have for that? Most notably memcg and
> > their high limit?
> 
> 
> Let me take a stab at this. The specific reasons why high limit is not a
> good interface to implement proactive reclaim:
> 
> 1) It can cause allocations from the target application to get
> throttled.
> 
> 2) It leaves a state (high limit) in the kernel which needs to be reset
> by the userspace part of proactive reclaimer.
> 
> If I remember correctly, Facebook actually tried to use high limit to
> implement the proactive reclaim but due to exactly these limitations [1]
> they went the route [2] aligned with this proposal.
> 
> To further explain why the above limitations are pretty bad: The
> proactive reclaimers usually use feedback loop to decide how much to
> squeeze from the target applications without impacting their performance
> or impacting within a tolerable range. The metrics used for the feedback
> loop are either refaults or PSI and these metrics becomes messy due to
> application getting throttled due to high limit.
> 
> For (2), the high limit interface is a very awkward interface to use to
> do proactive reclaim. If the userspace proactive reclaimer fails/crashed
> due to whatever reason during triggering the reclaim in an application,
> it can leave the application in a bad state (memory pressure state and
> throttled) for a long time.

Yes.

In addition to the proactive reclaimer crashing, we also had problems
of it simply not responding quickly enough.

Because there is a delay between reclaim (action) and refaults
(feedback), there is a very real upper limit of pages you can
reasonably reclaim per second, without risking pressure spikes that
far exceed tolerances. A fixed memory.high limit can easily exceed
that safe reclaim rate when the workload expands abruptly. Even if the
proactive reclaimer process is alive, it's almost impossible to step
between a rapidly allocating process and its cgroup limit in time.

The semantics of writing to memory.high also require that the new
limit is met before returning to userspace. This can take a long time,
during which the reclaimer cannot re-evaluate the optimal target size
based on observed pressure. We routinely saw the reclaimer get stuck
in the kernel hammering a suffering workload down to a stale target.

We tried for quite a while to make this work, but the limit semantics
turned out to not be a good fit for proactive reclaim.

A mechanism to request a fixed number of pages to reclaim turned out
to work much, much better in practice. We've been using a simple
per-cgroup knob (like here: https://lkml.org/lkml/2020/9/9/1094).

With tiered memory systems coming up, I can see the need for
restricting to specific numa nodes. Demoting from DRAM to CXL has a
different cost function than evicting RAM/CXL to storage, and those
two things probably need to happen at different rates.