On Thu 10-08-17 03:23:23, Daniel Colascione wrote:

Thanks for taking a look at the patch!

On Thu, Aug 10 2017, Michal Hocko wrote:

quoted

[CC linux-api - the patch was posted here
http://lkml.kernel.org/r/20170810001557.147285-1-dancol@google.com]

On Thu 10-08-17 13:38:31, Minchan Kim wrote:

quoted

On Wed, Aug 09, 2017 at 05:15:57PM -0700, Daniel Colascione wrote:

quoted

/proc/pid/smaps_rollup is a new proc file that improves the
performance of user programs that determine aggregate memory
statistics (e.g., total PSS) of a process.

Android regularly "samples" the memory usage of various processes in
order to balance its memory pool sizes. This sampling process involves
opening /proc/pid/smaps and summing certain fields. For very large
processes, sampling memory use this way can take several hundred
milliseconds, due mostly to the overhead of the seq_printf calls in
task_mmu.c.

Have you tried to reduce that overhead? E.g. by replacing seq_printf by
something more simple
http://lkml.kernel.org/r/20160817130320.GC20703@dhcp22.suse.cz?

I haven't tried that yet, but if I'm reading that thread correctly, it
looks like using more efficient printing primitives gives us a 7%
speedup. The smaps_rollup patch gives us a much bigger speedup while
reusing almost all the smaps code, so it seems easier and simpler than a
bunch of incremental improvements to smaps. And even an efficient smaps
would have to push 2MB through seq_file for the 3000-VMA process case.

The thing is that more users would benefit from a more efficient
/proc/pid/smaps call. Maybe we can use some caching tricks etc...  We
should make sure that existing options should be attempted before a new
user visible interface is added. It is kind of sad that the real work
(pte walk) is less expensive than formating the output and copying it to
the userspace...

quoted

How often you you need to read this information?

It varies depending on how often processes change state.  We sample a
short time (tens of seconds) after processes change state (e.g., enters
foreground) and every few minutes thereafter. We're particularly
concerned from an energy perspective about needlessly burning CPU on
background samples.

Please make sure this is documented in the patch along with some numbers
ideally.

[...]

quoted

quoted

FYI, there was trial but got failed at that time so in this time,
https://marc.info/?l=linux-kernel&m=147310650003277&w=2
http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1229163.html

Yes I really disliked the previous attempt and this one is not all that
better. The primary unanswered question back then was a relevant
usecase. Back then it was argued [1] that PSS was useful for userspace
OOM handling but arguments were rather dubious. Follow up questions [2]
shown that the useage of PSS was very workload specific. Minchan has
noted some usecase as well but not very specific either.

Anyway, I see what you mean about PSS being iffy for user-space OOM
processing (because PSS doesn't tell you how much memory you get back in
exchange for killing a given process at a particular moment). We're not
using it like that.

Instead, we're using the PSS samples we collect asynchronously for
system-management tasks like fine-tuning oom_adj_score, memory use
tracking for debugging, application-level memory-use attribution, and
deciding whether we want to kill large processes during system idle
maintenance windows. Android has been using PSS for these purposes for a
long time; as the average process VMA count has increased and and
devices become more efficiency-conscious, PSS-collection inefficiency
has started to matter more. IMHO, it'd be a lot safer to optimize the
existing PSS-collection model, which has been fine-tuned over the years,
instead of changing the memory tracking approach entirely to work around
smaps-generation inefficiency.

This is really vague. Please be more specific.
-- 
Michal Hocko
SUSE Labs