Re: [PATCH] mm: memcg: optimize parent iteration in memcg_rstat_updated()

[Yosry Ahmed <yosryahmed@google.com>]

On Wed, Jan 24, 2024 at 9:38 AM Shakeel Butt <shakeelb@google.com> wrote:
>
> On Wed, Jan 24, 2024 at 2:00 AM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > In memcg_rstat_updated(), we iterate the memcg being updated and its
> > parents to update memcg->vmstats_percpu->stats_updates in the fast path
> > (i.e. no atomic updates). According to my math, this is 3 memory loads
> > (and potentially 3 cache misses) per memcg:
> > - Load the address of memcg->vmstats_percpu.
> > - Load vmstats_percpu->stats_updates (based on some percpu calculation).
> > - Load the address of the parent memcg.
> >
> > Avoid most of the cache misses by caching a pointer from each struct
> > memcg_vmstats_percpu to its parent on the corresponding CPU. In this
> > case, for the first memcg we have 2 memory loads (same as above):
> > - Load the address of memcg->vmstats_percpu.
> > - Load vmstats_percpu->stats_updates (based on some percpu calculation).
> >
> > Then for each additional memcg, we need a single load to get the
> > parent's stats_updates directly. This reduces the number of loads from
> > O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.

This is actually O(1+N) not O(2+N). Every memcg needs one load, and
the first one needs an extra load.

> >
> > Additionally, stash a pointer to memcg->vmstats in each struct
> > memcg_vmstats_percpu such that we can access the atomic counter that all
> > CPUs fold into, memcg->vmstats->stats_updates.
> > memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
> > accept a struct memcg_vmstats pointer accordingly.
> >
> > In struct memcg_vmstats_percpu, make sure both pointers together with
> > stats_updates live on the same cacheline. Finally, update
> > mem_cgroup_alloc() to take in a parent pointer and initialize the new
> > cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
> > look concerning, but there are multiple similar loops in the cgroup
> > creation path (e.g. cgroup_rstat_init()), most of which are hidden
> > within alloc_percpu().
> >
> > According to Oliver's testing [1], this fixes multiple 30-38%
> > regressions in vm-scalability, will-it-scale-tlb_flush2, and
> > will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
> > CPU (<2KB on a machine with 128 CPUs).
> >
> > [1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
> >
> > Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
> > Tested-by: kernel test robot <oliver.sang@intel.com>
> > Reported-by: kernel test robot <oliver.sang@intel.com>
> > Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
> > Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
> > ---
>
> Nice work.
>
> Acked-by: Shakeel Butt <shakeelb@google.com>

Thanks!