Re: [RFC][PATCH 2/5] memcg: uncharge in batched manner

["KAMEZAWA Hiroyuki" <kamezawa.hiroyu@jp.fujitsu.com>]

Balbir Singh さんは書きました：
> * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-08-28
> 13:24:38]:
>
>>
>> In massive parallel enviroment, res_counter can be a performance
>> bottleneck.
>> This patch is a trial for reducing lock contention.
>> One strong techinque to reduce lock contention is reducing calls by
>> batching some amount of calls int one.
>>
>> Considering charge/uncharge chatacteristic,
>> 	- charge is done one by one via demand-paging.
>> 	- uncharge is done by
>> 		- in chunk at munmap, truncate, exit, execve...
>> 		- one by one via vmscan/paging.
>>
>> It seems we hace a chance to batched-uncharge.
>> This patch is a base patch for batched uncharge. For avoiding
>> scattering memcg's structure, this patch adds memcg batch uncharge
>> information to the task. please see start/end usage in next patch.
>>
>
> Overall it is a very good idea, can't we do the uncharge at the poin
> tof unmap_vmas, exit_mmap, etc so that we don't have to keep
> additional data structures around.
>
We can't. We uncharge when page->mapcount goes down to 0.
This is unknown until page_remove_rmap() decrement page->mapcount
by "atomic" ops.

My first version allocated memcg_batch_info on stack ...and..
I had to pass an extra argument to page_remove_rmap() etc....
That was very ugly ;(
Now, I adds per-task memcg_batch_info to task struct.
Because it will be always used at exit() and make exit() path
much faster, it's not very costly.

Thanks,
-Kame


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>