Re: [PATCH v2 6/6] zswap: memcg accounting

[Johannes Weiner <hannes-druUgvl0LCNAfugRpC6u6w@public.gmane.org>]

Hello Michal,

On Fri, May 13, 2022 at 05:14:26PM +0200, Michal Koutný wrote:
> On Wed, May 11, 2022 at 03:06:56PM -0400, Johannes Weiner <hannes-druUgvl0LCNAfugRpC6u6w@public.gmane.org> wrote:
> > Correct. After which the uncompressed page is reclaimed and uncharged.
> > So the zswapout process will reduce the charge bottom line.
> 
> A zswap object falling under memory.current was my first thinking, I was
> confused why it's exported as a separate counter memory.zswap.current
> (which IMO suggests disjoint counting) and it doubles a
> memory.stat:zswap entry.
> 
> Is the separate memory.zswap.current good for anything? (Except maybe
> avoiding global rstat flush on memory.stat read but that'd be an
> undesired precendent.)

Right, it's accounted as a subset rather than fully disjointed. But it
is a limitable counter of its own, so I exported it as such, with a
current and a max knob. This is comparable to the kmem counter in v1.

From an API POV it would be quite strange to have max for a counter
that has no current. Likewise it would be strange for a major memory
consumer to be missing from memory.stat.

> (Ad the eventually reduced footprint, the transitional excursion above
> memcg's (or ancestor's) limit should be limited by number of parallel
> reclaims running (each one at most a page, right?), so it doesn't seem
> necessary to tackle (now).)

Correct.

> > memory.zswap.* are there to configure zswap policy, within the
> > boundaries of available memory - it's by definition a subset.
> 
> I see how the .max works when equal to 0 or "max". The intermediate
> values are more difficult to reason about.

It needs to be configured to the workload's access frequency curve,
which can be done with trial-and-error (reasonable balance between
zswpins and pswpins) or in a more targeted manner using tools such as
page_idle, damon etc.

> Also, I can see that on the global level, zswap is configured relatively
> (/sys/module/zswap/parameters/max_pool_percent).
> You wrote that the actual configured value is workload specific, would
> it be simpler to have also relative zswap limit per memcg?
>
> (Relative wrt memory.max, it'd be rather just a convenience with this
> simple ratio, however, it'd correspond to the top level limit. OTOH, the
> relatives would have counter-intuitive hierarchical behavior. I don't
> mean this should be changed, rather wondering why this variant was
> chosen.)

A percentage isn't a bad way to pick a global default limit for a
kernel feature. But it would have been preferable if zswap had used
the percentage internally and made the knob based in bytes (like
min_free_kbytes for example).

Because for load tuning, bytes make much more sense. That's how you
measure the workingset, so a percentage is an awkward indirection. At
the cgroup level, it makes even less sense: all memcg tunables are in
bytes, it would be quite weird to introduce a "max" that is 0-100. Add
the confusion of how percentages would propagate down the hierarchy...

> > +bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
> > +{
> > +     struct mem_cgroup *memcg, *original_memcg;
> > +     bool ret = true;
> > +
> > +     original_memcg = get_mem_cgroup_from_objcg(objcg);
> > +     for (memcg = original_memcg; memcg != root_mem_cgroup;
> > +          memcg = parent_mem_cgroup(memcg)) {
> > +             unsigned long max = READ_ONCE(memcg->zswap_max);
> > +             unsigned long pages;
> > +
> > +             if (max == PAGE_COUNTER_MAX)
> > +                     continue;
> > +             if (max == 0) {
> > +                     ret = false;
> > +                     break;
> > +             }
> > +
> > +             cgroup_rstat_flush(memcg->css.cgroup);
> 
> Here, I think it'd be better not to bypass mem_cgroup_flush_stats() (the
> mechanism is approximate and you traverse all ancestors anyway), i.e.
> mem_cgroup_flush_stats() before the loop instead of this.

I don't traverse all ancestors, I bail on disabled groups and skip
unlimited ones. This saves a lot of flushes in practice right now: our
heaviest swapping cgroups have zswap disabled (max=0) because they're
lowpri and forced to disk. Likewise, the zswap users have their zswap
limit several levels down from the root, and I currently don't ever
flush the higher levels (max=PAGE_COUNTER_MAX).

Flushing unnecessary groups with a ratelimit doesn't sound like an
improvement to me.

Thanks