Re: [RFC][PATCH 0/7] memcg async reclaim

[Ying Han <yinghan@xxxxxxxxxx>]

Here I ran some tests and the result.

On a 32G machine, I created a memcg with 4G hard_limit (limit_in_bytes) and and ran cat on a 20g file. Then I use getdelays to measure the ttfp "delay average" under RECLAIM. When the workload is reaching its hard_limit and

without background reclaim, each ttfp is triggered by a pagefault. I would like to demostrate the average delay average for ttfp (thus page fault latency) on the streaming read/write workload and compare it w/ per-memcg bg reclaim enabled.

Note:

1. I applied a patch on getdelays.c from fengguang which shows average CPU/IO/SWAP/RECLAIM delays in ns.

2. I used my latest version of per-memcg-per-kswapd patch for the following test. The patch could have been improved since then and I can run the same test when Kame has his patch ready.

Configuration:

$ cat /proc/meminfo

MemTotal:       33045832 kB

$ cat /dev/cgroup/memory/A/memory.limit_in_bytes

4294967296

$ cat /dev/cgroup/memory/A/memory.reclaim_wmarks

low_wmark 4137680896

high_wmark 4085252096

Test:

$ echo $$ >/dev/cgroup/memory/A/tasks

$ cat /export/hdc3/dd_A/tf0 > /dev/zero

Without per-memcg background reclaim:

CPU             count     real total  virtual total    delay total  delay average

               176589    17248377848    27344548685     1093693318       6193.440ns

IO              count    delay total  delay average

               160704   242072632962        1506326ns

SWAP            count    delay total  delay average

                    0              0              0ns

RECLAIM         count    delay total  delay average

                15944     3512140153         220279ns

cat: read=20947877888, write=0, cancelled_write=0

real>---4m26.912s

user>---0m0.227s

sys>----0m27.823s

With per-memcg background reclaim:

$ ps -ef | grep memcg

root      5803     2  2 13:56 ?        00:04:20 [memcg_4]

CPU             count     real total  virtual total    delay total  delay average

               161085    13185995424    23863858944       72902585        452.572ns

IO              count    delay total  delay average

               160915   246145533109        1529661ns

SWAP            count    delay total  delay average

                    0              0              0ns

RECLAIM         count    delay total  delay average

                    0              0              0ns

cat: read=20974891008, write=0, cancelled_write=0

real>---4m26.572s

user>---0m0.246s

sys>----0m24.192s

memcg_4 cputime: 2.86sec

Observation:

1. Without the background reclaim, the cat hit ttfp heavely and the "delay average" goes above 220 microsec.

2. With background reclaim, the ttfp delay average is always 0. Since the ttfp happens synchronously and that implies the latency of the application overtime.

3. The real time goes slighly better w/ bg reclaim and the sys time is about the same ( adding the memcg_4 time on top of sys time of cat). But i don't expect big cpu benefit. The async reclaim uses spare cputime to proactivly reclaim pages on the side which gurantees less latency variation of application over time.

--Ying

On Thu, May 12, 2011 at 10:10 PM, Ying Han <yinghan@xxxxxxxxxx> wrote:

On Thu, May 12, 2011 at 8:03 PM, KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:

On Thu, 12 May 2011 17:17:25 +0900

KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:

> On Thu, 12 May 2011 13:22:37 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:

> I'll check what codes in vmscan.c or /mm affects memcg and post a
> required fix in step by step. I think I found some..
>

After some tests, I doubt that 'automatic' one is unnecessary until
memcg's dirty_ratio is supported. And as Andrew pointed out,
total cpu consumption is unchanged and I don't have workloads which
shows me meaningful speed up.

The total cpu consumption is one way to measure the background reclaim, another thing I would like to measure is a histogram of page fault latency

for a heavy page allocation application. I would expect with background reclaim, we will get less variation on the page fault latency than w/o it. 

Sorry i haven't got chance to run some tests to back it up. I will try to get some data.

 

But I guess...with dirty_ratio, amount of dirty pages in memcg is
limited and background reclaim can work enough without noise of
write_page() while applications are throttled by dirty_ratio.

Definitely. I have run into the issue while debugging the soft_limit reclaim. The background reclaim became very inefficient if we have dirty pages greater than the soft_limit. Talking w/ Greg about it regarding his per-memcg dirty page limit effort, we should consider setting the dirty ratio which not allowing the dirty pages greater the reclaim watermarks (here is the soft_limit).

--Ying

 

Hmm, I'll study for a while but it seems better to start active soft limit,
(or some threshold users can set) first.

Anyway, this work makes me to see vmscan.c carefully and I think I can
post some patches for fix, tunes.

Thanks,
-Kame