lkml.org 
[lkml]   [2012]   [Jul]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 00/34] Memory management performance backports for -stable
Date
This series is related to the new addition to stable_kernel_rules.txt

- Serious issues as reported by a user of a distribution kernel may also
be considered if they fix a notable performance or interactivity issue.
As these fixes are not as obvious and have a higher risk of a subtle
regression they should only be submitted by a distribution kernel
maintainer and include an addendum linking to a bugzilla entry if it
exists and additional information on the user-visible impact.

All of these patches have been backported to a distribution kernel and
address some sort of performance issue in the VM. As they are not all
obvious, I've added a "Stable note" to the top of each patch giving
additional information on why the patch was backported. Lets see where
the boundaries lie on how this new rule is interpreted in practice :).

Patch 1 Performance fix for tmpfs
Patch 2 Memory hotadd fix
Patch 3 Reduce boot time on large machines
Patches 4-5 Reduce stalls for wait_iff_congested
Patches 6-8 Reduce excessive reclaim of slab objects which for some workloads
will reduce the amount of IO required
Patches 9-10 limits the amount of page reclaim when THP/Compaction is active.
Excessive reclaim in low memory situations can lead to stalls some
of which are user visible.
Patches 11-19 reduce the amount of churn of the LRU lists. Poor reclaim
decisions can impair workloads in different ways and there have
been complaints recently the reclaim decisions of modern kernels
are worse than older ones.
Patches 20-21 reduce the amount of CPU kswapd uses in some cases. This
is harder to trigger but were developed due to bug reports about
100% CPU usage from kswapd.
Patches 22-25 are mostly related to interactivity when THP is enabled.
Patches 26-30 are also related to page reclaim decisions, particularly
the residency of mapped pages.
Patches 31-34 fix a major page allocator performance regression

All of the patches will apply to 3.0-stable but the ordering of the
patches is such that applying them to 3.2-stable and 3.4-stable should
be straight-forward.

I am bending or breaking the rules in places that needs examination.

1. Not all these patches have a bugzilla entry because in many cases I was
doing the investigation based on my own testing. By rights, I should
have been creating bugzilla entries for each of them but there only are
so many hours in the day.
2. I will be duplicated in the signed-offs because I may both the author
of the patch and now part of the submission path to -stable. I don't
think there is anything wrong with this but it might look weird to
some people.
3. Some patches are in the series only because they make later patches
easier to backport.
4. Patch 30 stomps all over the rules. The upstream patch accidentally
fixes a problem and was found through bisection but the full patch
and the series itself is not a good -stable candidate.

I'm running tests against the backport as it's a unique combination but the
patches have been tested as part of the distribution backport already. It'll
be a few days before I have an exact comparison between 3.0.36 and the
backport but I have a few basic results against 3.0.23. I'm not going to
analyse them in detail but here a few points

http://www.csn.ul.ie/~mel/postings/stableport-20120719/global-dhp__pagealloc-performance/hydra/comparison.html

o System CPU time reduced on kernbench
o Page allocator latency reduced for the most part, there are
counter-examples but it's mostly reduced
o page_test, brk_test improved on aim9
o 5% gain in page faults/sec in page fault micro benchmark

http://www.csn.ul.ie/~mel/postings/stableport-20120719/global-dhp__pagereclaim-performance-ext4/hydra/comparison.html
o fsmark in single threaded mode completed faster and with higher operations/second
o postmark looks slower but there are changes in ext4 between 3.0.23 and 3.0.36 that
might account for this. kswapd scan rates were slightly reduced
o In the micro benchmark, it took longer to complete but kswapd and direct
reclaim activity were reduced

Alex,Shi (2):
kswapd: avoid unnecessary rebalance after an unsuccessful balancing
kswapd: assign new_order and new_classzone_idx after wakeup in
sleeping

Dave Chinner (3):
vmscan: add shrink_slab tracepoints
vmscan: shrinker->nr updates race and go wrong
vmscan: reduce wind up shrinker->nr when shrinker can't do work

David Rientjes (2):
cpusets: avoid looping when storing to mems_allowed if one node
remains set
cpusets: stall when updating mems_allowed for mempolicy or disjoint
nodemask

Dimitri Sivanich (1):
mm: vmstat: cache align vm_stat

Hugh Dickins (1):
mm: test PageSwapBacked in lumpy reclaim

Johannes Weiner (1):
mm: vmscan: fix force-scanning small targets without swap

Konstantin Khlebnikov (3):
vmscan: promote shared file mapped pages
vmscan: activate executable pages after first usage
mm/hugetlb: fix warning in alloc_huge_page/dequeue_huge_page_vma

Mel Gorman (14):
mm: memory hotplug: Check if pages are correctly reserved on a
per-section basis
mm: Reduce the amount of work done when updating min_free_kbytes
mm: Abort reclaim/compaction if compaction can proceed
mm: migration: clean up unmap_and_move()
mm: compaction: Allow compaction to isolate dirty pages
mm: compaction: Determine if dirty pages can be migrated without
blocking within ->migratepage
mm: page allocator: Do not call direct reclaim for THP allocations
while compaction is deferred
mm: compaction: make isolate_lru_page() filter-aware again
mm: compaction: Introduce sync-light migration for use by compaction
mm: vmscan: When reclaiming for compaction, ensure there are
sufficient free pages available
mm: vmscan: Do not OOM if aborting reclaim to start compaction
mm: vmscan: Check if reclaim should really abort even if
compaction_ready() is true for one zone
mm: vmscan: Do not force kswapd to scan small targets
cpuset: mm: Reduce large amounts of memory barrier related damage v3

Minchan Kim (5):
mm: compaction: trivial clean up in acct_isolated()
mm: change isolate mode from #define to bitwise type
mm: compaction: make isolate_lru_page() filter-aware
mm: zone_reclaim: make isolate_lru_page() filter-aware
mm/vmscan.c: consider swap space when deciding whether to continue
reclaim

Rik van Riel (1):
mm: limit direct reclaim for higher order allocations

Shaohua Li (1):
vmscan: clear ZONE_CONGESTED for zone with good watermark

.../trace/postprocess/trace-vmscan-postprocess.pl | 8 +-
drivers/base/memory.c | 58 ++--
fs/btrfs/disk-io.c | 5 +-
fs/hugetlbfs/inode.c | 3 +-
fs/nfs/internal.h | 2 +-
fs/nfs/write.c | 4 +-
include/linux/cpuset.h | 45 ++--
include/linux/fs.h | 11 +-
include/linux/init_task.h | 8 +
include/linux/memcontrol.h | 3 +-
include/linux/migrate.h | 23 +-
include/linux/mmzone.h | 14 +
include/linux/sched.h | 2 +-
include/linux/swap.h | 7 +-
include/trace/events/vmscan.h | 85 +++++-
kernel/cpuset.c | 63 ++---
kernel/fork.c | 3 +
mm/compaction.c | 26 +-
mm/filemap.c | 11 +-
mm/hugetlb.c | 13 +-
mm/memcontrol.c | 3 +-
mm/memory-failure.c | 2 +-
mm/memory_hotplug.c | 2 +-
mm/mempolicy.c | 30 ++-
mm/migrate.c | 224 ++++++++++------
mm/page_alloc.c | 113 +++++---
mm/slab.c | 13 +-
mm/slub.c | 39 ++-
mm/vmscan.c | 280 ++++++++++++++++----
mm/vmstat.c | 2 +-
30 files changed, 772 insertions(+), 330 deletions(-)
--
1.7.9.2



\
 
 \ /
  Last update: 2012-07-19 17:41    [from the cache]
©2003-2011 Jasper Spaans