lkml.org 
[lkml]   [2018]   [Jun]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [LKP] [lkp-robot] [mm, memcontrol] 309fe96bfc: vm-scalability.throughput +23.0% improvement
On Fri, Jun 01, 2018 at 03:26:04PM +0800, Aaron Lu wrote:
> On Mon, May 28, 2018 at 07:40:19PM +0800, kernel test robot wrote:
> >
> > Greeting,
> >
> > FYI, we noticed a +23.0% improvement of vm-scalability.throughput due to commit:
> >
> >
> > commit: 309fe96bfc0ae387f53612927a8f0dc3eb056efd ("mm, memcontrol: implement memory.swap.events")
> > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
> >
> > in testcase: vm-scalability
> > on test machine: 144 threads Intel(R) Xeon(R) CPU E7-8890 v3 @ 2.50GHz with 512G memory
> > with following parameters:
> >
> > runtime: 300s
> > size: 1T
> > test: lru-shm
> > cpufreq_governor: performance
> >
> > test-description: The motivation behind this suite is to exercise functions and regions of the mm/ of the Linux kernel which are of interest to us.
> > test-url: https://git.kernel.org/cgit/linux/kernel/git/wfg/vm-scalability.git/
> >
>
> With the patch I just sent out:
> "mem_cgroup: make sure moving_account, move_lock_task and stat_cpu in the
> same cacheline"
>
> Applying this commit on top doesn't yield 23% improvement any more, but
> a 6% performace drop...
> I found the culprit being the following one line introduced in this commit:
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index d90b0201a8c4..07ab974c0a49 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -6019,13 +6019,17 @@ int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
> if (!memcg)
> return 0;
>
> - if (!entry.val)
> + if (!entry.val) {
> + memcg_memory_event(memcg, MEMCG_SWAP_FAIL);

Removing this line restored performance but it really doesn't make any
sense. Ying suggested it might be code alignment related and suggested
to use a different compiler than gcc-7.2. Then I used gcc-6.4 and turned
out the test result to be pretty much the same for the two commits:

(each test has run for 3 times)
$ grep throughput base/*/stats.json
base/0/stats.json: "vm-scalability.throughput": 89207489,
base/1/stats.json: "vm-scalability.throughput": 89982933,
base/2/stats.json: "vm-scalability.throughput": 90436592,

$ grep throughput head/*/stats.json
head/0/stats.json: "vm-scalability.throughput": 90882775,
head/1/stats.json: "vm-scalability.throughput": 90675220,
head/2/stats.json: "vm-scalability.throughput": 91173479,

So probably it's really related to code alignment and this bisected
commit doesn't cause performance change(as expected).

> return 0;
> + }
>
> memcg = mem_cgroup_id_get_online(memcg);
>
> If I remove that memcg_memory_event() call, performance will restore.
>
> It's beyond my understanding why this code path matters since there is
> no swap device setup in the test machine so I don't see how possible
> get_swap_page() could ever be called.
>
> Still investigating...
>

\
 
 \ /
  Last update: 2018-06-06 10:51    [W:0.062 / U:41.628 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site