Re: [PATCH V2 0/6] Per-cgroup page stat accounting

[Michal Hocko <mhocko@suse.cz>]

On Fri 27-07-12 18:20:32, Sha Zhengju wrote:
> From: Sha Zhengju <handai.szj@taobao.com>
> 
> Hi, list
> 
> This V2 patch series provide the ability for each memory cgroup to
> have independent dirty/writeback page statistics which can provide
> information for per-cgroup direct reclaim or some.
>
> In the first three prepare patches, we have done some cleanup and
> reworked vfs set page dirty routines to make "modify page info" and
> "dirty page accouting" stay in one function as much as possible for
> the sake of memcg bigger lock(test numbers are in the specific patch).

I think the first 3 patches should be merged (if VFS people are OK with
them of course) first to make the rest of the work easier.

> Kame, I tested these patches on linux mainline v3.5, because I cannot
> boot up the kernel under linux-next :(. But these patches are cooked
> on top of your recent memcg patches (I backport them to mainline) and
> I think there is no hunk with the mm tree.  So If there's no other
> problem, I think it could be considered for merging.
> 
> 
> 
> Following is performance comparison between before/after the series:
> 
> Test steps(Mem-24g, ext4):
> drop_cache; sync
> cat /proc/meminfo|grep Dirty (=4KB)
> fio (buffered/randwrite/bs=4k/size=128m/filesize=1g/numjobs=8/sync) 
> cat /proc/meminfo|grep Dirty(=648696kB)

How does this tests that the accounting is correct? I would expect
something that would exercise dirty vs. writeback and having 0 in the
end. with your 24G memcg and 1G IO you do not see any memcg effects
(only the global ones). Have you tested with something like 20M group and
pushing 1G data through it?

> 
> We test it for 10 times and get the average numbers:
> Before:
> write: io=1024.0MB, bw=334678 KB/s, iops=83669.2 , runt=  3136 msec
> lat (usec): min=1 , max=26203.1 , avg=81.473, stdev=275.754
> 
> After:
> write: io=1024.0MB, bw=325219 KB/s, iops= 81304.1 , runt=  3226.9 msec
> lat (usec): min=1 , max=17224 , avg=86.194, stdev=298.183
> 
> 
> 
> There is about 2.8% performance decrease. But I notice that once memcg
> is enabled, the root_memcg exsits and all pages allocated are belong
> to it, so they will go through the root memcg statistics routines
> which bring some overhead.

Yes, this is exactly what I had in mind when I asked about the overhead
when memcg are not in use last time (mentioning jump labels or how they
are called at the moment).

> Moreover,in case of memcg_is_enable && no cgroups, we can get root
> memcg stats just from global numbers which can avoid both accounting
> overheads and many if-test overheads. Later I'll work further into it.

I think it is a must before merging. We do not want an infrastructure
that is not used currently and which brings an overhead.
Anyway, this is a good start for the future work. Thanks for doing that!

> Any comments are welcomed. : )
> 
> 
> 
> Change log:
> v2 <-- v1:
> 	1. add test numbers
> 	2. some small fix and comments
> 
> Sha Zhengju (6):
> 	memcg-remove-MEMCG_NR_FILE_MAPPED.patch
> 	Make-TestSetPageDirty-and-dirty-page-accounting-in-o.patch
> 	Use-vfs-__set_page_dirty-interface-instead-of-doing-.patch
> 	memcg-add-per-cgroup-dirty-pages-accounting.patch
> 	memcg-add-per-cgroup-writeback-pages-accounting.patch
> 	memcg-Document-cgroup-dirty-writeback-memory-statist.patch
> 
>  Documentation/cgroups/memory.txt |    2 +
>  fs/buffer.c                      |   36 +++++++++++++++--------
>  fs/ceph/addr.c                   |   20 +------------
>  include/linux/buffer_head.h      |    2 +
>  include/linux/memcontrol.h       |   30 ++++++++++++++-----
>  mm/filemap.c                     |    9 ++++++
>  mm/memcontrol.c                  |   58 +++++++++++++++++++-------------------
>  mm/page-writeback.c              |   48 ++++++++++++++++++++++++++++---
>  mm/rmap.c                        |    4 +-
>  mm/truncate.c                    |    6 ++++
>  10 files changed, 141 insertions(+), 74 deletions(-)
> 
> --
> 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>

-- 
Michal Hocko
SUSE Labs

--
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>