lkml.org 
[lkml]   [2012]   [Jun]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v4 00/25] kmem limitation for memcg
(2012/06/18 19:27), Glauber Costa wrote:
> Hello All,
>
> This is my new take for the memcg kmem accounting. This should merge
> all of the previous comments from you guys, specially concerning the big churn
> inside the allocators themselves.
>
> My focus in this new round was to keep the changes in the cache internals to
> a minimum. To do that, I relied upon two main pillars:
>
> * Cristoph's unification series, that allowed me to put must of the changes
> in a common file. Even then, the changes are not too many, since the overal
> level of invasiveness was decreased.
> * Accounting is done directly from the page allocator. This means some pages
> can fail to be accounted, but that can only happen when the task calling
> kmem_cache_alloc or kmalloc is not the same task allocating a new page.
> This never happens in steady state operation if the tasks are kept in the
> same memcg. Naturally, if the page ends up being accounted to a memcg that
> is not limited (such as root memcg), that particular page will simply not
> be accounted.
>
> The dispatcher code stays (mem_cgroup_get_kmem_cache), being the mechanism who
> guarantees that, during steady state operation, all objects allocated in a page
> will belong to the same memcg. I consider this a good compromise point between
> strict and loose accounting here.
>

2 questions.

- Do you have performance numbers ?

- Do you think user-memory memcg should be switched to page-allocator level accounting ?
(it will require some study for modifying current bached-freeing and per-cpu-stock
logics...)

Thanks,
-Kame



\
 
 \ /
  Last update: 2012-06-18 15:41    [W:0.299 / U:2.852 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site