On Thu 22-02-18 18:38:11, Andrey Ryabinin wrote: > > > On 02/22/2018 06:33 PM, Michal Hocko wrote: > > On Thu 22-02-18 18:13:11, Andrey Ryabinin wrote: > >> > >> > >> On 02/22/2018 05:09 PM, Michal Hocko wrote: > >>> On Thu 22-02-18 16:50:33, Andrey Ryabinin wrote: > >>>> On 02/21/2018 11:17 PM, Andrew Morton wrote: > >>>>> On Fri, 19 Jan 2018 16:11:18 +0100 Michal Hocko <mhocko@xxxxxxxxxx> wrote: > >>>>> > >>>>>> And to be honest, I do not really see why keeping retrying from > >>>>>> mem_cgroup_resize_limit should be so much faster than keep retrying from > >>>>>> the direct reclaim path. We are doing SWAP_CLUSTER_MAX batches anyway. > >>>>>> mem_cgroup_resize_limit loop adds _some_ overhead but I am not really > >>>>>> sure why it should be that large. > >>>>> > >>>>> Maybe restarting the scan lots of times results in rescanning lots of > >>>>> ineligible pages at the start of the list before doing useful work? > >>>>> > >>>>> Andrey, are you able to determine where all that CPU time is being spent? > >>>>> > >>>> > >>>> I should have been more specific about the test I did. The full script looks like this: > >>>> > >>>> mkdir -p /sys/fs/cgroup/memory/test > >>>> echo $$ > /sys/fs/cgroup/memory/test/tasks > >>>> cat 4G_file > /dev/null > >>>> while true; do cat 4G_file > /dev/null; done & > >>>> loop_pid=$! > >>>> perf stat echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes > >>>> echo -1 > /sys/fs/cgroup/memory/test/memory.limit_in_bytes > >>>> kill $loop_pid > >>>> > >>>> > >>>> I think the additional loops add some overhead and it's not that big by itself, but > >>>> this small overhead allows task to refill slightly more pages, increasing > >>>> the total amount of pages that mem_cgroup_resize_limit() need to reclaim. > >>>> > >>>> By using the following commands to show the the amount of reclaimed pages: > >>>> perf record -e vmscan:mm_vmscan_memcg_reclaim_end echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes > >>>> perf script|cut -d '=' -f 2| paste -sd+ |bc > >>>> > >>>> I've got 1259841 pages (4.9G) with the patch vs 1394312 pages (5.4G) without it. > >>> > >>> So how does the picture changes if you have multiple producers? > >>> > >> > >> Drastically, in favor of the patch. But numbers *very* fickle from run to run. > >> > >> Inside 5G vm with 4 cpus (qemu -m 5G -smp 4) and 4 processes in cgroup reading 1G files: > >> "while true; do cat /1g_f$i > /dev/null; done &" > >> > >> with the patch: > >> best: 1.04 secs, 9.7G reclaimed > >> worst: 2.2 secs, 16G reclaimed. > >> > >> without: > >> best: 5.4 sec, 35G reclaimed > >> worst: 22.2 sec, 136G reclaimed > > > > Could you also compare how much memory do we reclaim with/without the > > patch? > > > > I did and I wrote the results. Please look again. I must have forgotten. Care to point me to the message-id? 20180119132544.19569-2-aryabinin@xxxxxxxxxxxxx doesn't contain this information and a quick glance over the follow up thread doesn't have anything as well. Ideally, this should be in the patch changelog, btw. -- Michal Hocko SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
