Re: [PATCH RFC v2 4/4] memcg: implement memory thresholds

[KAMEZAWA Hiroyuki]

On Tue, 15 Dec 2009 12:46:32 +0200
"Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:

> On Tue, Dec 15, 2009 at 3:58 AM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > On Sat, 12 Dec 2009 00:59:19 +0200
> > "Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:

> > If you use have to use spinlock here, this is a system-wide spinlock,
> > threshold as "100" is too small, I think.
> 
> What is reasonable value for THRESHOLDS_EVENTS_THRESH for you?
> 
> In most cases spinlock taken only for two checks. Is it significant time?
> 
I tend to think about "bad case" when I see spinlock. 

And...I'm not sure but, recently, there are many VM users.
spinlock can be a big pitfall in some enviroment if not para-virtualized.
(I'm sorry I misunderstand somehing and VM handle this well...)

> Unfortunately, I can't test it on a big box. I have only dual-core system.
> It's not enough to test scalability.
> 

please leave it as 100 for now. But there is a chance to do simple optimization
for reducing the number of checks.

example)
static void mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
{
	/* For handle memory allocation in rush, check jiffies */
	*/
	smp_rmb();
	if (memcg->last_checkpoint_jiffies == jiffies)
		return;   /* reset event to half value ..*/
	memcg->last_checkpoint_jiffies = jiffies;
	smp_wmb();
	.....

I think this kind of check is necessary for handle "Rushing" memory allocation
in scalable way. Above one is just an example, 1 tick may be too long.

Other simple plan is

	/* Allow only one thread to do scan the list at the same time. */
	if (atomic_inc_not_zero(&memcg->threahold_scan_count) {
		atomic_dec(&memcg->threshold_scan_count);
		return;
	}
	...
	atomic_dec(&memcg->threahold_scan_count)

Some easy logic (as above) for taking care of scalability and commenary for that
is enough at 1st stage. Then, if there seems to be a trouble/concern, someone
(me?) will do some work later.




Thanks,
-Kame

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/