On Tue, Dec 20, 2022 at 11:53:25AM -0800, Shakeel Butt wrote:
> +Vlastimil
>
> On Tue, Dec 20, 2022 at 10:48 AM Roman Gushchin
> <roman.gushchin@xxxxxxxxx> wrote:
> >
> > Sven Luther reported a regression in the posix message queues
> > performance caused by switching to the per-object tracking of
> > slab objects introduced by patch series ending with the
> > commit 10befea91b61 ("mm: memcg/slab: use a single set of kmem_caches for all
> > allocations").
> >
> > To mitigate the regression cache allocated mqueue messages on a small
> > percpu cache instead of releasing and re-allocating them every time.
> >
>
> Seems fine with me but I am wondering what is stopping us to do this
> caching in the slab layer for all accounted allocations? Does this
> only make sense for specific scenarios/use-cases?
It's far from trivial, unfortunately. Here we have an mqueue object to associate
a percpu cache with and the hit rate is expected to be high, assuming the mqueue
will be used to pass a lot of messages.
With a generic slab cache we return to the necessity of managing
the per-cgroup x per-slab-cache x per-cpu free list (or some other data structure),
which is already far from trivial, based on the previous experience. It can
easily lead to a significant memory waste, which will fully compensate all perf
wins.
So probably we need some heuristics to allocate caches only for really hot slab
caches and use some sort of a hash map (keyed by cgroup and slab cache) to
access freelists. What I miss to commit more time to this project (aside from not
having it), is the lack of real workloads which will noticeably win from this work.
Sven provided a good example and benchmark to reproduce the regression, so it
was easy to justify the work.
Thanks!
