Re: [PATCH v3 3/5] mm/memcg: Cache vmstat data in percpu memcg_stock_pcp

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

On Thu, Apr 15, 2021 at 01:08:29PM -0400, Waiman Long wrote:
> On 4/15/21 12:50 PM, Johannes Weiner wrote:
> > On Tue, Apr 13, 2021 at 09:20:25PM -0400, Waiman Long wrote:
> > > Before the new slab memory controller with per object byte charging,
> > > charging and vmstat data update happen only when new slab pages are
> > > allocated or freed. Now they are done with every kmem_cache_alloc()
> > > and kmem_cache_free(). This causes additional overhead for workloads
> > > that generate a lot of alloc and free calls.
> > > 
> > > The memcg_stock_pcp is used to cache byte charge for a specific
> > > obj_cgroup to reduce that overhead. To further reducing it, this patch
> > > makes the vmstat data cached in the memcg_stock_pcp structure as well
> > > until it accumulates a page size worth of update or when other cached
> > > data change.
> > > 
> > > On a 2-socket Cascade Lake server with instrumentation enabled and this
> > > patch applied, it was found that about 17% (946796 out of 5515184) of the
> > > time when __mod_obj_stock_state() is called leads to an actual call to
> > > mod_objcg_state() after initial boot. When doing parallel kernel build,
> > > the figure was about 16% (21894614 out of 139780628). So caching the
> > > vmstat data reduces the number of calls to mod_objcg_state() by more
> > > than 80%.
> > Right, but mod_objcg_state() is itself already percpu-cached. What's
> > the benefit of avoiding calls to it with another percpu cache?
> > 
> There are actually 2 set of vmstat data that have to be updated. One is
> associated with the memcg and other one is for each lruvec within the
> cgroup. Caching it in obj_stock, we replace 2 writes to two colder
> cachelines with one write to a hot cacheline. If you look at patch 5, I
> break obj_stock into two - one for task context and one for irq context.
> Interrupt disable is no longer needed in task context, but that is not
> possible when writing to the actual vmstat data arrays.

Ah, thanks for the explanation. Both of these points are worth
mentioning in the changelog of this patch.