Re: [PATCH 1/4] per-task-delay-accounting: add memory reclaim delay

[Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>]

KOSAKI Motohiro wrote:
>>>> +	delayacct_freepages_start();
>>>>  	did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
>>>> +	delayacct_freepages_end();
>>> Shouldn't we consider memcgroup reclaim?
>> thanks for pointing this.
>> Unfortunately, we're not so familiar to memcgroup.
>> Will look into this.
>> Could you tell us pointers to memcgroup?
> 
> Documentation/controllers/memory.txt is good guide :)

thanks, and playing with this is good too.
I see it's easy to use this feature :)

> 
> <abstract>
> try_to_free_pages():	         global reclaim entry point
> try_to_free_mem_cgroup_pages():  memcgroup reclaim entry point
> do_try_to_free_pages():	         common layer

I see the point that, on memcgroup enabled system, there are
memcgroup memory reclaim points and try_to_free_mem_cgroup_pages()
is called when the page charge reach the limit.
So if we want to account delay for memory reclaim, we should
account at both of try_to_free_pages() and try_to_free_mem_cgroup_pages().
Accounting at do_try_to_free_pages() covers this case.

> Unfortunately, I don't know your requirement and
> memcgroup reclaim shold be mesured.
> I hope you explain your purpose and benefit more.

Ah, sure.
We have an issue that a process which accesses a DB sometime slows down.
We want to know what causes it, without the reproducer.
We have only the report of the problem, memory usage, vmstat...
And, I don't know the memory reclaim really causes the problem.

If we can see what kind of delay is the matter in the test bed, it will
help us to take a next action.

Thanks,
Hiroshi Shimamoto