Re: [LKP] Re: [mm/memcg] bd0b230fe1: will-it-scale.per_process_ops -22.7% regression

[Feng Tang]

On Fri, Nov 20, 2020 at 07:44:24PM +0800, Feng Tang wrote:
> On Fri, Nov 13, 2020 at 03:34:36PM +0800, Feng Tang wrote:
> > > I would rather focus on a more effective mem_cgroup layout. It is very
> > > likely that we are just stumbling over two counters here.
> > > 
> > > Could you try to add cache alignment of counters after memory and see
> > > which one makes the difference? I do not expect memsw to be the one
> > > because that one is used together with the main counter. But who knows
> > > maybe the way it crosses the cache line has the exact effect. Hard to
> > > tell without other numbers.
> > 
> > I added some alignments change around the 'memsw', but neither of them can
> > restore the -22.7%. Following are some log showing how the alignments
> > are:
> > 
> > tl: memcg=0x7cd1000 memory=0x7cd10d0 memsw=0x7cd1140 kmem=0x7cd11b0 tcpmem=0x7cd1220
> > t2: memcg=0x7cd0000 memory=0x7cd00d0 memsw=0x7cd0140 kmem=0x7cd01c0 tcpmem=0x7cd0230
> > 
> > So both of the 'memsw' are aligned, but t2's 'kmem' is aligned while
> > t1's is not.
> > 
> > I will check more on the perf data about detailed hotspots.
> 
> Some more check updates about it:
> 
> Waiman's patch is effectively removing one 'struct page_counter' between
> 'memory' and "memsw'. And the mem_cgroup is: 
> 
> struct mem_cgroup {
> 
> 	...
> 
> 	struct page_counter memory;		/* Both v1 & v2 */
> 
> 	union {
> 		struct page_counter swap;	/* v2 only */
> 		struct page_counter memsw;	/* v1 only */
> 	};
> 
> 	/* Legacy consumer-oriented counters */
> 	struct page_counter kmem;		/* v1 only */
> 	struct page_counter tcpmem;		/* v1 only */
> 
> 	...
> 	...
> 
> 	MEMCG_PADDING(_pad1_);
> 
> 	atomic_t		moving_account;
> 	struct task_struct	*move_lock_task;
> 	
> 	...
> };
> 
> 
> I do experiments by inserting a 'page_counter' between 'memory'
> and the 'MEMCG_PADDING(_pad1_)', no matter where I put it, the
> benchmark result can be recovered from 145K to 185K, which is
> really confusing, as adding a 'page_counter' right before the
> '_pad1_' doesn't change cache alignment of any members.

I think we finally found the trick :), further debugging shows it
is not related to the alignment inside one cacheline, but the
adjacency of 2 adjacent cacheliens (2N and 2N+1, one pair of 128 bytes).

For structure mem_cgroup, member 'vmstats_local', 'vmstats_percpu'
sit in one cacheline, while 'vmstats[]' sits in the next cacheline,
and when 'adjacent cacheline prefetch" is enabled, if these 2 lines
sit in one pair (128 btyes), say 2N and 2N+1, then there seems to
be some kind of false sharing, and if they sit in 2 pairs, say
2N-1 and 2N then it's fine.

And with the following patch to relayout these members, the regression
is restored and event better. while reducing 64 bytes of sizeof
'struct mem_cgroup'

	parent_commit	Waiman's_commit		+relayout patch

result	187K		145K			200K

Also, if we disable the hw prefetch feature, the Waiman's commit
and its parent commit will have no performance difference.

Thanks,
Feng