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.
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.
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.
quoted
quoted
smaps_rollup improves the situation. It contains most of the fields of
/proc/pid/smaps, but instead of a set of fields for each VMA,
smaps_rollup instead contains one synthetic smaps-format entry
representing the whole process. In the single smaps_rollup synthetic
entry, each field is the summation of the corresponding field in all
of the real-smaps VMAs. Using a common format for smaps_rollup and
smaps allows userspace parsers to repurpose parsers meant for use with
non-rollup smaps for smaps_rollup, and it allows userspace to switch
between smaps_rollup and smaps at runtime (say, based on the
availability of smaps_rollup in a given kernel) with minimal fuss.
By using smaps_rollup instead of smaps, a caller can avoid the
significant overhead of formatting, reading, and parsing each of a
large process's potentially very numerous memory mappings. For
sampling system_server's PSS in Android, we measured a 12x speedup,
representing a savings of several hundred milliseconds.
By a large process you mean a process with many VMAs right? How many
vmas are we talking about?
Yes: ~3000 VMAs is one I'd consider large in this context.
quoted
quoted
One alternative to a new per-process proc file would have been
including PSS information in /proc/pid/status. We considered this
option but thought that PSS would be too expensive (by a few orders of
magnitude) to collect relative to what's already emitted as part of
/proc/pid/status, and slowing every user of /proc/pid/status for the
sake of readers that happen to want PSS feels wrong.
The code itself works by reusing the existing VMA-walking framework we
use for regular smaps generation and keeping the mem_size_stats
structure around between VMA walks instead of using a fresh one for
each VMA. In this way, summation happens automatically. We let
seq_file walk over the VMAs just as it does for regular smaps and just
emit nothing to the seq_file until we hit the last VMA.
Patch changelog:
v2: Fix typo in commit message
Add ABI documentation as requested by gregkh
Signed-off-by: Daniel Colascione <redacted>
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.
The existence of an independent attempt to add this functionality
suggests that it might be generally useful too.
So let's start with a clear use case description. Then let's make it
clear that even optimizing the current implementation is not sufficient
to meat goals and only then try to add one more user visible API which
we will have to maintain for ever.
Adding a new API shouldn't be anyone's first choice, but the efficiency
wins are hard to ignore. We can cut ~2MB of smaps data to a few KB this
way. I'm definitely open to other ideas for getting similar wins, but
right now, I'd like to explore approaches that let us keep using the
existing PSS metric.
Thanks again!
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
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.
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.
[...]
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
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.
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
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.
[...]
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.
I actually think this is really similar to the Chrome OS use case --
we need to do proper accounting of memory from user space, and we need
something more accurate than what we have now (usually RSS) to figure
it out. I'm not sure what is vague about that statement?
PSS is not perfect but in closed systems where we have some knowledge
about what is being shared amongst process, PSS is much better than
RSS and readily available. So, I disagree that this is a dubious
usage -- if there's a better metric for making this kind of decision,
please share it.
Also I realized there's another argument for presenting this
information outside of smaps which is that we expose far less
information about a process and it's address space via something like
this, so it's much better for isolation to have a separate file with
different permissions. Right now the process in charge of accounting
for memory usage also gains knowledge about each process's address
space which is unnecessary.
IMHO, the fact that multiple folks have independently asked for this
seems like an argument that something like this is needed.
From: Tim Murray <hidden> Date: 2017-08-10 19:17:07
I've looked into this a fair bit on the Android side, so I can provide
some context. There are two main reasons why Android gathers PSS
information:
1. Android devices can show the user the amount of memory used per
application via the settings app. This is a less important use case.
2. We log PSS to help identify leaks in applications. We have found an
enormous number of bugs (in the Android platform, in Google's own
apps, and in third-party applications) using this data.
To do this, system_server (the main process in Android userspace) will
sample the PSS of a process three seconds after it changes state (for
example, app is launched and becomes the foreground application) and
about every ten minutes after that. The net result is that PSS
collection is regularly running on at least one process in the system
(usually a few times a minute while the screen is on, less when screen
is off due to suspend). PSS of a process is an incredibly useful stat
to track, and we aren't going to get rid of it. We've looked at some
very hacky approaches using RSS ("take the RSS of the target process,
subtract the RSS of the zygote process that is the parent of all
Android apps") to reduce the accounting time, but it regularly
overestimated the memory used by 20+ percent. Accordingly, I don't
think that there's a good alternative to using PSS.
We started looking into PSS collection performance after we noticed
random frequency spikes while a phone's screen was off; occasionally,
one of the CPU clusters would ramp to a high frequency because there
was 200-300ms of constant CPU work from a single thread in the main
Android userspace process. The work causing the spike (which is
reasonable governor behavior given the amount of CPU time needed) was
always PSS collection. As a result, Android is burning more power than
we should be on PSS collection.
The other issue (and why I'm less sure about improving smaps as a
long-term solution) is that the number of VMAs per process has
increased significantly from release to release. After trying to
figure out why we were seeing these 200-300ms PSS collection times on
Android O but had not noticed it in previous versions, we found that
the number of VMAs in the main system process increased by 50% from
Android N to Android O (from ~1800 to ~2700) and varying increases in
every userspace process. Android M to N also had an increase in the
number of VMAs, although not as much. I'm not sure why this is
increasing so much over time, but thinking about ASLR and ways to make
ASLR better, I expect that this will continue to increase going
forward. I would not be surprised if we hit 5000 VMAs on the main
Android process (system_server) by 2020.
If we assume that the number of VMAs is going to increase over time,
then doing anything we can do to reduce the overhead of each VMA
during PSS collection seems like the right way to go, and that means
outputting an aggregate statistic (to avoid whatever overhead there is
per line in writing smaps and in reading each line from userspace).
Also, Dan sent me some numbers from his benchmark measuring PSS on
system_server (the big Android process) using smaps vs smaps_rollup:
using smaps:
iterations:1000 pid:1163 pss:220023808
0m29.46s real 0m08.28s user 0m20.98s system
using smaps_rollup:
iterations:1000 pid:1163 pss:220702720
0m04.39s real 0m00.03s user 0m04.31s system
On Thu, Aug 10, 2017 at 11:56 AM, Sonny Rao [off-list ref] wrote:
On Thu, Aug 10, 2017 at 3:58 AM, Michal Hocko [off-list ref] wrote:
quoted
On Thu 10-08-17 03:23:23, Daniel Colascione wrote:
quoted
Thanks for taking a look at the patch!
On Thu, Aug 10 2017, Michal Hocko wrote:
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.
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
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.
[...]
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.
I actually think this is really similar to the Chrome OS use case --
we need to do proper accounting of memory from user space, and we need
something more accurate than what we have now (usually RSS) to figure
it out. I'm not sure what is vague about that statement?
PSS is not perfect but in closed systems where we have some knowledge
about what is being shared amongst process, PSS is much better than
RSS and readily available. So, I disagree that this is a dubious
usage -- if there's a better metric for making this kind of decision,
please share it.
Also I realized there's another argument for presenting this
information outside of smaps which is that we expose far less
information about a process and it's address space via something like
this, so it's much better for isolation to have a separate file with
different permissions. Right now the process in charge of accounting
for memory usage also gains knowledge about each process's address
space which is unnecessary.
IMHO, the fact that multiple folks have independently asked for this
seems like an argument that something like this is needed.
From: Michal Hocko <mhocko@kernel.org> Date: 2017-08-24 08:55:53
Sorry for a late reply
On Thu 10-08-17 12:17:07, Tim Murray wrote:
I've looked into this a fair bit on the Android side, so I can provide
some context. There are two main reasons why Android gathers PSS
information:
1. Android devices can show the user the amount of memory used per
application via the settings app. This is a less important use case.
yes
2. We log PSS to help identify leaks in applications. We have found an
enormous number of bugs (in the Android platform, in Google's own
apps, and in third-party applications) using this data.
To do this, system_server (the main process in Android userspace) will
sample the PSS of a process three seconds after it changes state (for
example, app is launched and becomes the foreground application) and
about every ten minutes after that. The net result is that PSS
collection is regularly running on at least one process in the system
(usually a few times a minute while the screen is on, less when screen
is off due to suspend). PSS of a process is an incredibly useful stat
to track, and we aren't going to get rid of it. We've looked at some
very hacky approaches using RSS ("take the RSS of the target process,
subtract the RSS of the zygote process that is the parent of all
Android apps") to reduce the accounting time, but it regularly
overestimated the memory used by 20+ percent. Accordingly, I don't
think that there's a good alternative to using PSS.
Even if the RSS overestimates this shouldn't hide a memory leak, no?
We started looking into PSS collection performance after we noticed
random frequency spikes while a phone's screen was off; occasionally,
one of the CPU clusters would ramp to a high frequency because there
was 200-300ms of constant CPU work from a single thread in the main
Android userspace process. The work causing the spike (which is
reasonable governor behavior given the amount of CPU time needed) was
always PSS collection. As a result, Android is burning more power than
we should be on PSS collection.
Yes, this really sucks but we are revolving around the same point. It
really sucks that we burn so much time just copying the output to the
userspace when the real stuff (vma walk and pte walk) has to be done
anyway. AFAIR I could reduce the overhead by using more appropriate
seq_* functions but maybe we can do even better.
The other issue (and why I'm less sure about improving smaps as a
long-term solution) is that the number of VMAs per process has
increased significantly from release to release. After trying to
figure out why we were seeing these 200-300ms PSS collection times on
Android O but had not noticed it in previous versions, we found that
the number of VMAs in the main system process increased by 50% from
Android N to Android O (from ~1800 to ~2700) and varying increases in
every userspace process. Android M to N also had an increase in the
number of VMAs, although not as much. I'm not sure why this is
increasing so much over time, but thinking about ASLR and ways to make
ASLR better, I expect that this will continue to increase going
forward. I would not be surprised if we hit 5000 VMAs on the main
Android process (system_server) by 2020.
The thing is, however, that the larger amount of VMAs will also mean
more work on the kernel side. The data collection has to be done anyway.
If we assume that the number of VMAs is going to increase over time,
then doing anything we can do to reduce the overhead of each VMA
during PSS collection seems like the right way to go, and that means
outputting an aggregate statistic (to avoid whatever overhead there is
per line in writing smaps and in reading each line from userspace).
Also, Dan sent me some numbers from his benchmark measuring PSS on
system_server (the big Android process) using smaps vs smaps_rollup:
using smaps:
iterations:1000 pid:1163 pss:220023808
0m29.46s real 0m08.28s user 0m20.98s system
using smaps_rollup:
iterations:1000 pid:1163 pss:220702720
0m04.39s real 0m00.03s user 0m04.31s system
I would assume we would do all we can to reduce this kernel->user
overhead first before considering a new user visible file. I haven't
seen any attempts except from the low hanging fruid I have tried.
--
Michal Hocko
SUSE Labs
From: Andrew Morton <akpm@linux-foundation.org> Date: 2017-08-25 21:16:37
On Thu, 24 Aug 2017 10:55:53 +0200 Michal Hocko [off-list ref] wrote:
quoted
If we assume that the number of VMAs is going to increase over time,
then doing anything we can do to reduce the overhead of each VMA
during PSS collection seems like the right way to go, and that means
outputting an aggregate statistic (to avoid whatever overhead there is
per line in writing smaps and in reading each line from userspace).
Also, Dan sent me some numbers from his benchmark measuring PSS on
system_server (the big Android process) using smaps vs smaps_rollup:
using smaps:
iterations:1000 pid:1163 pss:220023808
0m29.46s real 0m08.28s user 0m20.98s system
using smaps_rollup:
iterations:1000 pid:1163 pss:220702720
0m04.39s real 0m00.03s user 0m04.31s system
I would assume we would do all we can to reduce this kernel->user
overhead first before considering a new user visible file. I haven't
seen any attempts except from the low hanging fruid I have tried.
It's hard to believe that we'll get anything like a 5x speedup via
optimization of the existing code?
From: Michal Hocko <mhocko@kernel.org> Date: 2017-08-28 11:30:23
On Fri 25-08-17 14:16:37, Andrew Morton wrote:
On Thu, 24 Aug 2017 10:55:53 +0200 Michal Hocko [off-list ref] wrote:
quoted
quoted
If we assume that the number of VMAs is going to increase over time,
then doing anything we can do to reduce the overhead of each VMA
during PSS collection seems like the right way to go, and that means
outputting an aggregate statistic (to avoid whatever overhead there is
per line in writing smaps and in reading each line from userspace).
Also, Dan sent me some numbers from his benchmark measuring PSS on
system_server (the big Android process) using smaps vs smaps_rollup:
using smaps:
iterations:1000 pid:1163 pss:220023808
0m29.46s real 0m08.28s user 0m20.98s system
using smaps_rollup:
iterations:1000 pid:1163 pss:220702720
0m04.39s real 0m00.03s user 0m04.31s system
I would assume we would do all we can to reduce this kernel->user
overhead first before considering a new user visible file. I haven't
seen any attempts except from the low hanging fruid I have tried.
It's hard to believe that we'll get anything like a 5x speedup via
optimization of the existing code?
Maybe we will not get that much of a boost but having misleading numbers
really quick is not something we should aim for. Just try to think what
the cumulative numbers actually mean. How can you even consider
cumulative PSS when you have no idea about mappings that were
considered?
--
Michal Hocko
SUSE Labs