Re: [RFC][PATCH v3 0/10] memcg async reclaim

[KAMEZAWA Hiroyuki]

On Thu, 26 May 2011 18:49:26 -0700
Ying Han <yinghan@xxxxxxxxxx> wrote:

> On Wed, May 25, 2011 at 10:10 PM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> >
> > It's now merge window...I just dump my patch queue to hear other's idea.
> > I wonder I should wait until dirty_ratio for memcg is queued to mmotm...
> > I'll be busy with LinuxCon Japan etc...in the next week.
> >
> > This patch is onto mmotm-May-11 + some patches queued in mmotm, as numa_stat.
> >
> > This is a patch for memcg to keep margin to the limit in background.
> > By keeping some margin to the limit in background, application can
> > avoid foreground memory reclaim at charge() and this will help latency.
> >
> > Main changes from v2 is.
> > Â- use SCHED_IDLE.
> > Â- removed most of heuristic codes. Now, code is very simple.
> >
> > By using SCHED_IDLE, async memory reclaim can only consume 0.3%? of cpu
> > if the system is truely busy but can use much CPU if the cpu is idle.
> > Because my purpose is for reducing latency without affecting other running
> > applications, SCHED_IDLE fits this work.
> >
> > If application need to stop by some I/O or event, background memory reclaim
> > will cull memory while the system is idle.
> >
> > Perforemce:
> > ÂRunning an httpd (apache) under 300M limit. And access 600MB working set
> > Âwith normalized distribution access by apatch-bench.
> > Âapatch bench's concurrency was 4 and did 40960 accesses.
> >
> > Without async reclaim:
> > Connection Times (ms)
> >       Âmin Âmean[+/-sd] median  max
> > Connect: Â Â Â Â0 Â Â0 Â 0.0 Â Â Â0 Â Â Â 2
> > Processing: Â Â30 Â 37 Â28.3 Â Â 32 Â Â1793
> > Waiting: Â Â Â 28 Â 35 Â25.5 Â Â 31 Â Â1792
> > Total: Â Â Â Â 30 Â 37 Â28.4 Â Â 32 Â Â1793
> >
> > Percentage of the requests served within a certain time (ms)
> > Â50% Â Â 32
> > Â66% Â Â 32
> > Â75% Â Â 33
> > Â80% Â Â 34
> > Â90% Â Â 39
> > Â95% Â Â 60
> > Â98% Â Â100
> > Â99% Â Â133
> > Â100% Â 1793 (longest request)
> >
> > With async reclaim:
> > Connection Times (ms)
> >       Âmin Âmean[+/-sd] median  max
> > Connect: Â Â Â Â0 Â Â0 Â 0.0 Â Â Â0 Â Â Â 2
> > Processing: Â Â30 Â 35 Â12.3 Â Â 32 Â Â 678
> > Waiting: Â Â Â 28 Â 34 Â12.0 Â Â 31 Â Â 658
> > Total: Â Â Â Â 30 Â 35 Â12.3 Â Â 32 Â Â 678
> >
> > Percentage of the requests served within a certain time (ms)
> > Â50% Â Â 32
> > Â66% Â Â 32
> > Â75% Â Â 33
> > Â80% Â Â 34
> > Â90% Â Â 39
> > Â95% Â Â 49
> > Â98% Â Â 71
> > Â99% Â Â 86
> > Â100% Â Â678 (longest request)
> >
> >
> > It seems latency is stabilized by hiding memory reclaim.
> >
> > The score for memory reclaim was following.
> > See patch 10 for meaning of each member.
> >
> > == without async reclaim ==
> > recent_scan_success_ratio 44
> > limit_scan_pages 388463
> > limit_freed_pages 162238
> > limit_elapsed_ns 13852159231
> > soft_scan_pages 0
> > soft_freed_pages 0
> > soft_elapsed_ns 0
> > margin_scan_pages 0
> > margin_freed_pages 0
> > margin_elapsed_ns 0
> >
> > == with async reclaim ==
> > recent_scan_success_ratio 6
> > limit_scan_pages 0
> > limit_freed_pages 0
> > limit_elapsed_ns 0
> > soft_scan_pages 0
> > soft_freed_pages 0
> > soft_elapsed_ns 0
> > margin_scan_pages 1295556
> > margin_freed_pages 122450
> > margin_elapsed_ns 644881521
> >
> >
> > For this case, SCHED_IDLE workqueue can reclaim enough memory to the httpd.
> >
> > I may need to dig why scan_success_ratio is far different in the both case.
> > I guess the difference of epalsed_ns is because several threads enter
> > memory reclaim when async reclaim doesn't run. But may not...
> >
> 
> 
> Hmm.. I noticed a very strange behavior on a simple test w/ the patch set.
> 
> Test:
> I created a 4g memcg and start doing cat. Then the memcg being OOM
> killed as soon as it reaches its hard_limit. We shouldn't hit OOM even
> w/o async-reclaim.
> 
> Again, I will read through the patch. But like to post the test result first.
> 
> $ echo $$ >/dev/cgroup/memory/A/tasks
> $ cat /dev/cgroup/memory/A/memory.limit_in_bytes
> 4294967296
> 
> $ time cat /export/hdc3/dd_A/tf0 > /dev/zero
> Killed
> 

I did the same kind of test without any problem...but ok, I'll do more test
later.



> real	0m53.565s
> user	0m0.061s
> sys	0m4.814s
> 
> Here is the OOM log:
> 
> May 26 18:43:00  kernel: [  963.489112] cat invoked oom-killer:
> gfp_mask=0xd0, order=0, oom_adj=0, oom_score_adj=0
> May 26 18:43:00  kernel: [  963.489121] Pid: 9425, comm: cat Tainted:
> G        W   2.6.39-mcg-DEV #131
> May 26 18:43:00  kernel: [  963.489123] Call Trace:
> May 26 18:43:00  kernel: [  963.489134]  [<ffffffff810e3512>]
> dump_header+0x82/0x1af
> May 26 18:43:00  kernel: [  963.489137]  [<ffffffff810e33ca>] ?
> spin_lock+0xe/0x10
> May 26 18:43:00  kernel: [  963.489140]  [<ffffffff810e33f9>] ?
> find_lock_task_mm+0x2d/0x67
> May 26 18:43:00  kernel: [  963.489143]  [<ffffffff810e38dd>]
> oom_kill_process+0x50/0x27b
> May 26 18:43:00  kernel: [  963.489155]  [<ffffffff810e3dc6>]
> mem_cgroup_out_of_memory+0x9a/0xe4
> May 26 18:43:00  kernel: [  963.489160]  [<ffffffff811153aa>]
> mem_cgroup_handle_oom+0x134/0x1fe
> May 26 18:43:00  kernel: [  963.489163]  [<ffffffff81114a72>] ?
> __mem_cgroup_insert_exceeded+0x83/0x83
> May 26 18:43:00  kernel: [  963.489176]  [<ffffffff811166e9>]
> __mem_cgroup_try_charge.clone.3+0x368/0x43a
> May 26 18:43:00  kernel: [  963.489179]  [<ffffffff81117586>]
> mem_cgroup_cache_charge+0x95/0x123
> May 26 18:43:00  kernel: [  963.489183]  [<ffffffff810e16d8>]
> add_to_page_cache_locked+0x42/0x114
> May 26 18:43:00  kernel: [  963.489185]  [<ffffffff810e17db>]
> add_to_page_cache_lru+0x31/0x5f
> May 26 18:43:00  kernel: [  963.489189]  [<ffffffff81145636>]
> mpage_readpages+0xb6/0x132
> May 26 18:43:00  kernel: [  963.489194]  [<ffffffff8119992f>] ?
> noalloc_get_block_write+0x24/0x24
> May 26 18:43:00  kernel: [  963.489197]  [<ffffffff8119992f>] ?
> noalloc_get_block_write+0x24/0x24
> May 26 18:43:00  kernel: [  963.489201]  [<ffffffff81036742>] ?
> __switch_to+0x160/0x212
> May 26 18:43:00  kernel: [  963.489205]  [<ffffffff811978b2>]
> ext4_readpages+0x1d/0x1f
> May 26 18:43:00  kernel: [  963.489209]  [<ffffffff810e8d4b>]
> __do_page_cache_readahead+0x144/0x1e3
> May 26 18:43:00  kernel: [  963.489212]  [<ffffffff810e8e0b>]
> ra_submit+0x21/0x25
> May 26 18:43:00  kernel: [  963.489215]  [<ffffffff810e9075>]
> ondemand_readahead+0x18c/0x19f
> May 26 18:43:00  kernel: [  963.489218]  [<ffffffff810e9105>]
> page_cache_async_readahead+0x7d/0x86
> May 26 18:43:00  kernel: [  963.489221]  [<ffffffff810e2b7e>]
> generic_file_aio_read+0x2d8/0x5fe
> May 26 18:43:00  kernel: [  963.489225]  [<ffffffff81119626>]
> do_sync_read+0xcb/0x108
> May 26 18:43:00  kernel: [  963.489230]  [<ffffffff811f168a>] ?
> fsnotify_perm+0x66/0x72
> May 26 18:43:00  kernel: [  963.489233]  [<ffffffff811f16f7>] ?
> security_file_permission+0x2e/0x33
> May 26 18:43:00  kernel: [  963.489236]  [<ffffffff8111a0c8>]
> vfs_read+0xab/0x107
> May 26 18:43:00  kernel: [  963.489239]  [<ffffffff8111a1e4>] sys_read+0x4a/0x6e
> May 26 18:43:00  kernel: [  963.489244]  [<ffffffff8140f469>]
> sysenter_dispatch+0x7/0x27
> May 26 18:43:00  kernel: [  963.489248] Task in /A killed as a result
> of limit of /A
> May 26 18:43:00  kernel: [  963.489251] memory: usage 4194304kB, limit
> 4194304kB, failcnt 26
> May 26 18:43:00  kernel: [  963.489253] memory+swap: usage 0kB, limit
> 9007199254740991kB, failcnt 0
> 

Hmm, why memory+swap usage 0kb here...

In this set, I used mem_cgroup_margin() rather than res_counter_margin().
Hmm, do you disable swap accounting ? If so, I may miss some.

Thanks,
-Kame




--
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/