On Sun, May 19, 2019 at 8:53 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
>
> - Background
>
> The Android terminology used for forking a new process and starting an app
> from scratch is a cold start, while resuming an existing app is a hot start.
> While we continually try to improve the performance of cold starts, hot
> starts will always be significantly less power hungry as well as faster so
> we are trying to make hot start more likely than cold start.
>
> To increase hot start, Android userspace manages the order that apps should
> be killed in a process called ActivityManagerService. ActivityManagerService
> tracks every Android app or service that the user could be interacting with
> at any time and translates that into a ranked list for lmkd(low memory
> killer daemon). They are likely to be killed by lmkd if the system has to
> reclaim memory. In that sense they are similar to entries in any other cache.
> Those apps are kept alive for opportunistic performance improvements but
> those performance improvements will vary based on the memory requirements of
> individual workloads.
>
> - Problem
>
> Naturally, cached apps were dominant consumers of memory on the system.
> However, they were not significant consumers of swap even though they are
> good candidate for swap. Under investigation, swapping out only begins
> once the low zone watermark is hit and kswapd wakes up, but the overall
> allocation rate in the system might trip lmkd thresholds and cause a cached
> process to be killed(we measured performance swapping out vs. zapping the
> memory by killing a process. Unsurprisingly, zapping is 10x times faster
> even though we use zram which is much faster than real storage) so kill
> from lmkd will often satisfy the high zone watermark, resulting in very
> few pages actually being moved to swap.
It is not clear what exactly is the problem from the above para. IMO
low usage of swap is not the problem but rather global memory pressure
and the reactive response to it is the problem. Killing apps over swap
is preferred as you have noted zapping frees memory faster but it
indirectly increases cold start. Also swapping on allocation causes
latency issues for the app. So, a proactive mechanism is needed to
keep global pressure away and indirectly reduces cold starts and alloc
stalls.
>
> - Approach
>
> The approach we chose was to use a new interface to allow userspace to
> proactively reclaim entire processes by leveraging platform information.
> This allowed us to bypass the inaccuracy of the kernel’s LRUs for pages
> that are known to be cold from userspace and to avoid races with lmkd
> by reclaiming apps as soon as they entered the cached state. Additionally,
> it could provide many chances for platform to use much information to
> optimize memory efficiency.
I think it would be good to have clear reasoning on why "reclaim from
userspace" approach is taken. Android runtime clearly has more
accurate stale/cold information at the app/process level and can
positively influence kernel's reclaim decisions. So, "reclaim from
userspace" approach makes total sense for Android. I envision that
Chrome OS would be another very obvious user of this approach. There
can be tens of tabs which the user have not touched for sometime.
Chrome OS can proactively reclaim memory from such tabs.
>
> IMHO we should spell it out that this patchset complements MADV_WONTNEED
MADV_DONTNEED? same at couple of places below.
> and MADV_FREE by adding non-destructive ways to gain some free memory
> space. MADV_COLD is similar to MADV_WONTNEED in a way that it hints the
> kernel that memory region is not currently needed and should be reclaimed
> immediately; MADV_COOL is similar to MADV_FREE in a way that it hints the
> kernel that memory region is not currently needed and should be reclaimed
> when memory pressure rises.
>
> To achieve the goal, the patchset introduce two new options for madvise.
> One is MADV_COOL which will deactive activated pages and the other is
> MADV_COLD which will reclaim private pages instantly. These new options
> complement MADV_DONTNEED and MADV_FREE by adding non-destructive ways to
> gain some free memory space. MADV_COLD is similar to MADV_DONTNEED in a way
> that it hints the kernel that memory region is not currently needed and
> should be reclaimed immediately; MADV_COOL is similar to MADV_FREE in a way
> that it hints the kernel that memory region is not currently needed and
> should be reclaimed when memory pressure rises.
>
> This approach is similar in spirit to madvise(MADV_WONTNEED), but the
> information required to make the reclaim decision is not known to the app.
> Instead, it is known to a centralized userspace daemon, and that daemon
> must be able to initiate reclaim on its own without any app involvement.
> To solve the concern, this patch introduces new syscall -
>
> struct pr_madvise_param {
> int size;
> const struct iovec *vec;
> }
>
> int process_madvise(int pidfd, ssize_t nr_elem, int *behavior,
> struct pr_madvise_param *restuls,
> struct pr_madvise_param *ranges,
> unsigned long flags);
>
> The syscall get pidfd to give hints to external process and provides
> pair of result/ranges vector arguments so that it could give several
> hints to each address range all at once.
>
> I guess others have different ideas about the naming of syscall and options
> so feel free to suggest better naming.
>
> - Experiment
>
> We did bunch of testing with several hundreds of real users, not artificial
> benchmark on android. We saw about 17% cold start decreasement without any
> significant battery/app startup latency issues. And with artificial benchmark
> which launches and switching apps, we saw average 7% app launching improvement,
> 18% less lmkd kill and good stat from vmstat.
>
> A is vanilla and B is process_madvise.
>
>
> A B delta ratio(%)
> allocstall_dma 0 0 0 0.00
> allocstall_movable 1464 457 -1007 -69.00
> allocstall_normal 263210 190763 -72447 -28.00
> allocstall_total 264674 191220 -73454 -28.00
> compact_daemon_wake 26912 25294 -1618 -7.00
> compact_fail 17885 14151 -3734 -21.00
> compact_free_scanned 4204766409 3835994922 -368771487 -9.00
> compact_isolated 3446484 2967618 -478866 -14.00
> compact_migrate_scanned 1621336411 1324695710 -296640701 -19.00
> compact_stall 19387 15343 -4044 -21.00
> compact_success 1502 1192 -310 -21.00
> kswapd_high_wmark_hit_quickly 234 184 -50 -22.00
> kswapd_inodesteal 221635 233093 11458 5.00
> kswapd_low_wmark_hit_quickly 66065 54009 -12056 -19.00
> nr_dirtied 259934 296476 36542 14.00
> nr_vmscan_immediate_reclaim 2587 2356 -231 -9.00
> nr_vmscan_write 1274232 2661733 1387501 108.00
> nr_written 1514060 2937560 1423500 94.00
> pageoutrun 67561 55133 -12428 -19.00
> pgactivate 2335060 1984882 -350178 -15.00
> pgalloc_dma 13743011 14096463 353452 2.00
> pgalloc_movable 0 0 0 0.00
> pgalloc_normal 18742440 16802065 -1940375 -11.00
> pgalloc_total 32485451 30898528 -1586923 -5.00
> pgdeactivate 4262210 2930670 -1331540 -32.00
> pgfault 30812334 31085065 272731 0.00
> pgfree 33553970 31765164 -1788806 -6.00
> pginodesteal 33411 15084 -18327 -55.00
> pglazyfreed 0 0 0 0.00
> pgmajfault 551312 1508299 956987 173.00
> pgmigrate_fail 43927 29330 -14597 -34.00
> pgmigrate_success 1399851 1203922 -195929 -14.00
> pgpgin 24141776 19032156 -5109620 -22.00
> pgpgout 959344 1103316 143972 15.00
> pgpgoutclean 4639732 3765868 -873864 -19.00
> pgrefill 4884560 3006938 -1877622 -39.00
> pgrotated 37828 25897 -11931 -32.00
> pgscan_direct 1456037 957567 -498470 -35.00
> pgscan_direct_throttle 0 0 0 0.00
> pgscan_kswapd 6667767 5047360 -1620407 -25.00
> pgscan_total 8123804 6004927 -2118877 -27.00
> pgskip_dma 0 0 0 0.00
> pgskip_movable 0 0 0 0.00
> pgskip_normal 14907 25382 10475 70.00
> pgskip_total 14907 25382 10475 70.00
> pgsteal_direct 1118986 690215 -428771 -39.00
> pgsteal_kswapd 4750223 3657107 -1093116 -24.00
> pgsteal_total 5869209 4347322 -1521887 -26.00
> pswpin 417613 1392647 975034 233.00
> pswpout 1274224 2661731 1387507 108.00
> slabs_scanned 13686905 10807200 -2879705 -22.00
> workingset_activate 668966 569444 -99522 -15.00
> workingset_nodereclaim 38957 32621 -6336 -17.00
> workingset_refault 2816795 2179782 -637013 -23.00
> workingset_restore 294320 168601 -125719 -43.00
>
> pgmajfault is increased by 173% because swapin is increased by 200% by
> process_madvise hint. However, swap read based on zram is much cheaper
> than file IO in performance point of view and app hot start by swapin is
> also cheaper than cold start from the beginning of app which needs many IO
> from storage and initialization steps.
>
> This patchset is against on next-20190517.
>
> Minchan Kim (7):
> mm: introduce MADV_COOL
> mm: change PAGEREF_RECLAIM_CLEAN with PAGE_REFRECLAIM
> mm: introduce MADV_COLD
> mm: factor out madvise's core functionality
> mm: introduce external memory hinting API
> mm: extend process_madvise syscall to support vector arrary
> mm: madvise support MADV_ANONYMOUS_FILTER and MADV_FILE_FILTER
>
> arch/x86/entry/syscalls/syscall_32.tbl | 1 +
> arch/x86/entry/syscalls/syscall_64.tbl | 1 +
> include/linux/page-flags.h | 1 +
> include/linux/page_idle.h | 15 +
> include/linux/proc_fs.h | 1 +
> include/linux/swap.h | 2 +
> include/linux/syscalls.h | 2 +
> include/uapi/asm-generic/mman-common.h | 12 +
> include/uapi/asm-generic/unistd.h | 2 +
> kernel/signal.c | 2 +-
> kernel/sys_ni.c | 1 +
> mm/madvise.c | 600 +++++++++++++++++++++----
> mm/swap.c | 43 ++
> mm/vmscan.c | 80 +++-
> 14 files changed, 680 insertions(+), 83 deletions(-)
>
> --
> 2.21.0.1020.gf2820cf01a-goog
>
