Re: [PATCH 5/7] memcg: get rid of once-per-second cache shrinkingfor dead memcgs

[Glauber Costa]

On 11/16/2012 06:55 PM, Michal Hocko wrote:
> On Fri 16-11-12 16:21:59, KAMEZAWA Hiroyuki wrote:
>> (2012/11/16 16:11), Glauber Costa wrote:
>>> On 11/16/2012 09:07 AM, Kamezawa Hiroyuki wrote:
>>>> (2012/11/15 22:47), Glauber Costa wrote:
>>>>> On 11/15/2012 01:41 PM, Kamezawa Hiroyuki wrote:
>>>>>> (2012/11/15 11:54), Glauber Costa wrote:
>>>>>>> The idea is to synchronously do it, leaving it up to the shrinking
>>>>>>> facilities in vmscan.c and/or others. Not actively retrying shrinking
>>>>>>> may leave the caches alive for more time, but it will remove the ugly
>>>>>>> wakeups. One would argue that if the caches have free objects but are
>>>>>>> not being shrunk, it is because we don't need that memory yet.
>>>>>>>
>>>>>>> Signed-off-by: Glauber Costa <glommer@xxxxxxxxxxxxx>
>>>>>>> CC: Michal Hocko <mhocko@xxxxxxx>
>>>>>>> CC: Kamezawa Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
>>>>>>> CC: Johannes Weiner <hannes@xxxxxxxxxxx>
>>>>>>> CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
>>>>>>
>>>>>> I agree this patch but can we have a way to see the number of unaccounted
>>>>>> zombie cache usage for debugging ?
>>>>>>
>>>>>> Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
>>>>>>
>>>>> Any particular interface in mind ?
>>>>>
>>>>
>>>> Hmm, it's debug interface and having cgroup file may be bad.....
>>>> If it can be seen in bytes or some, /proc/vmstat ?
>>>>
>>>> out_of_track_slabs  xxxxxxx. hm ?
>>>>
>>>
>>> I particularly think that, being this a debug interface, it is also
>>> useful to have an indication of which caches are still in place. This is
>>> because the cache itself, is the best indication we have about the
>>> specific workload that may be keeping it in memory.
>>>
>>> I first thought debugfs could help us probing useful information out of
>>> it, but given all the abuse people inflicted in debugfs... maybe we
>>> could have a file in the root memcg with that information for all
>>> removed memcgs? If we do that, we can go further and list the memcgs
>>> that are pending due to memsw as well. memory.dangling_memcgs ?
>>>
>>
>> Hm, I'm ok with it... others ?
> 
> What about memory.kmem.dangling_caches?
> 
If that is what it does, sure.

But as I said, kmem is not the only thing that can keep caches in
memory. If we're going for this, maybe we should be more comprehensive
and show it all.

--
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/