On Fri, Nov 03, 2017 at 11:33:36AM -0400, Johannes Weiner wrote:
> We've seen memory.stat reads in top-level cgroups take up to fourteen
> seconds during a userspace bug that created tens of thousands of ghost
> cgroups pinned by lingering page cache.
>
> Even with a more reasonable number of cgroups, aggregating memory.stat
> is unnecessarily heavy. The complexity is this:
>
> nr_cgroups * nr_stat_items * nr_possible_cpus
>
> where the stat items are ~70 at this point. With 128 cgroups and 128
> CPUs - decent, not enormous setups - reading the top-level memory.stat
> has to aggregate over a million per-cpu counters. This doesn't scale.
>
> Instead of spreading the source of truth across all CPUs, use the
> per-cpu counters merely to batch updates to shared atomic counters.
>
> This is the same as the per-cpu stocks we use for charging memory to
> the shared atomic page_counters, and also the way the global vmstat
> counters are implemented.
>
> Vmstat has elaborate spilling thresholds that depend on the number of
> CPUs, amount of memory, and memory pressure - carefully balancing the
> cost of counter updates with the amount of per-cpu error. That's
> because the vmstat counters are system-wide, but also used for
> decisions inside the kernel (e.g. NR_FREE_PAGES in the
> allocator). Neither is true for the memory controller.
>
> Use the same static batch size we already use for page_counter updates
> during charging. The per-cpu error in the stats will be 128k, which is
> an acceptable ratio of cores to memory accounting granularity.
>
> Signed-off-by: Johannes Weiner <[email protected]>
> ---
> include/linux/memcontrol.h | 96 +++++++++++++++++++++++++++---------------
> mm/memcontrol.c | 101 +++++++++++++++++++++++----------------------
> 2 files changed, 113 insertions(+), 84 deletions(-)
Acked-by: Vladimir Davydov <[email protected]>
From 1583059644730153766@xxx Fri Nov 03 15:35:24 +0000 2017
X-GM-THRID: 1583059644730153766
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread