[PATCH 0/4] Reducing fragmentation using lists (sub-zones) v22

* [PATCH 0/4] Reducing fragmentation using lists (sub-zones) v22
@ 2006-01-20 11:54 Mel Gorman
  2006-01-20 11:54 ` [PATCH 1/4] Add __GFP_EASYRCLM flag and update callers Mel Gorman
                   ` (3 more replies)
  0 siblings, 4 replies; 11+ messages in thread
From: Mel Gorman @ 2006-01-20 11:54 UTC (permalink / raw)
  To: linux-mm; +Cc: jschopp, Mel Gorman, linux-kernel, kamezawa.hiroyu, lhms-devel

This is a rebase of the list-based anti-fragmentation approach to act as
a comparison to the zone-based approach. To recall, this version does *not*
depend on a usemap bitfield and instead uses a page flag to track where a page
should be freed to. Fundamentally, the list-based approach works by having
two free lists for each order - one for kernel pages and the other for user
pages. The per-cpu allocator is also split between kernel and user pages.

In total, it adds 146 lines of code, can be fully disabled via a compile-time
option without any performance penalty and when enabled, it was able allocate
about 70% of physical memory in 4MiB contiguous chunks after a heavy set of
benchmarks completed.

This gives best-effort for low fragmentation in all zones. Fully guaranteed
reclaim for hotplug and high order allocation in a zone would require a
zone. Full benchmarks are below the Changelog, but here is the summary

			clean		anti-frag
Kernel extract		14		14		(no difference)
Kernel build		741		741		(no difference)
aim9-page_test		131762.75	134311.90	+2549.15  (1.93%)
aim9-brk_test		586206.90	597167.14	+10960.24 (1.87%)
HugePages under load	60		86		+26     (+43%)
HugePages after tests	141		242		+101	(+71%)

The majorify of the aim9 results measured are within 1% of the standard
allocator. Kernel compiles vary sometimes by about 1-2 seconds or about
0.1%. The HugeTLB pages under load was 7 simulatanous kernel compiles,
each -j1.

When anti-frag is disabled via the config option, it behaves and performs just
like the standard allocator with no performance impact I can measure. When it
is enabled, there are definite variances but they are small and on loads like
kernel builds, it makes less than 3 seconds of a difference over 12 minutes.

Diffstat for full set of patches
 fs/buffer.c                |    3 
 fs/compat.c                |    2 
 fs/exec.c                  |    2 
 fs/inode.c                 |    2 
 include/asm-i386/page.h    |    3 
 include/linux/gfp.h        |    3 
 include/linux/highmem.h    |    3 
 include/linux/mmzone.h     |   28 ++++++-
 include/linux/page-flags.h |    7 +
 init/Kconfig               |   12 +++
 mm/memory.c                |    6 +
 mm/page_alloc.c            |  180 ++++++++++++++++++++++++++++++++++-----------
 mm/shmem.c                 |    4 -
 mm/swap_state.c            |    3 
 14 files changed, 202 insertions(+), 56 deletions(-)

Changelog since v21
o Update to 2.6.16-rc1-mm1

Changelog since v20
o Update to 2.6.15-rc1-mm2
o Remove usemap to reduce cache footprint, also does not hurt hotplug add
o Extend free_area not use new free_area lists, reduces footprint and is faster
o Leverage existing support for per-cpu page drain to avoid free but pinned
  pages
o Small number of micro-optimisations

Changelog since complex anti-fragmentation v19
o Updated to 2.6.15-rc1-mm2
o Removed the fallback area and balancing code
o Only differentiate between kernel and easy reclaimable allocations
  - Removes almost all the code that deals with usemaps
  - Made a number of simplifications based on two allocation lists
  - Fallback code is drastically simpler
o Do not change behaviour for high-order allocations
o Drop stats patch - unnecessary complications

Changelog since v18
o Resync against 2.6.14-rc5-mm1
o 004_markfree dropped
o Documentation note added on the behavior of free_area.nr_free

Changelog since v17
o Update to 2.6.14-rc4-mm1
o Remove explicit casts where implicit casts were in place
o Change __GFP_USER to __GFP_EASYRCLM, RCLM_USER to RCLM_EASY and PCPU_USER to
  PCPU_EASY
o Print a warning and return NULL if both RCLM flags are set in the GFP flags
o Reduce size of fallback_allocs
o Change magic number 64 to FREE_AREA_USEMAP_SIZE
o CodingStyle regressions cleanup
o Move sparsemen setup_usemap() out of header
o Changed fallback_balance to a mechanism that depended on zone->present_pages
  to avoid hotplug problems later
o Many superfluous parenthesis removed

Changlog since v16
o Variables using bit operations now are unsigned long. Note that when used
  as indices, they are integers and cast to unsigned long when necessary.
  This is because aim9 shows regressions when used as unsigned longs 
  throughout (~10% slowdown)
o 004_showfree added to provide more debugging information
o 008_stats dropped. Even with CONFIG_ALLOCSTATS disabled, it is causing 
  severe performance regressions. No explanation as to why
o for_each_rclmtype_order moved to header
o More coding style cleanups

Changelog since V14 (V15 not released)
o Update against 2.6.14-rc3
o Resync with Joel's work. All suggestions made on fix-ups to his last
  set of patches should also be in here. e.g. __GFP_USER is still __GFP_USER
  but is better commented.
o Large amount of CodingStyle, readability cleanups and corrections pointed
  out by Dave Hansen.
o Fix CONFIG_NUMA error that corrupted per-cpu lists
o Patches broken out to have one-feature-per-patch rather than
  more-code-per-patch
o Fix fallback bug where pages for RCLM_NORCLM end up on random other
  free lists.

Changelog since V13
o Patches are now broken out
o Added per-cpu draining of userrclm pages
o Brought the patch more in line with memory hotplug work
o Fine-grained use of the __GFP_USER and __GFP_KERNRCLM flags
o Many coding-style corrections
o Many whitespace-damage corrections

Changelog since V12
o Minor whitespace damage fixed as pointed by Joel Schopp

Changelog since V11
o Mainly a redefiff against 2.6.12-rc5
o Use #defines for indexing into pcpu lists
o Fix rounding error in the size of usemap

Changelog since V10
o All allocation types now use per-cpu caches like the standard allocator
o Removed all the additional buddy allocator statistic code
o Elimated three zone fields that can be lived without
o Simplified some loops
o Removed many unnecessary calculations

Changelog since V9
o Tightened what pools are used for fallbacks, less likely to fragment
o Many micro-optimisations to have the same performance as the standard 
  allocator. Modified allocator now faster than standard allocator using
  gcc 3.3.5
o Add counter for splits/coalescing

Changelog since V8
o rmqueue_bulk() allocates pages in large blocks and breaks it up into the
  requested size. Reduces the number of calls to __rmqueue()
o Beancounters are now a configurable option under "Kernel Hacking"
o Broke out some code into inline functions to be more Hotplug-friendly
o Increased the size of reserve for fallbacks from 10% to 12.5%. 

Changelog since V7
o Updated to 2.6.11-rc4
o Lots of cleanups, mainly related to beancounters
o Fixed up a miscalculation in the bitmap size as pointed out by Mike Kravetz
  (thanks Mike)
o Introduced a 10% reserve for fallbacks. Drastically reduces the number of
  kernnorclm allocations that go to the wrong places
o Don't trigger OOM when large allocations are involved

Changelog since V6
o Updated to 2.6.11-rc2
o Minor change to allow prezeroing to be a cleaner looking patch

Changelog since V5
o Fixed up gcc-2.95 errors
o Fixed up whitespace damage

Changelog since V4
o No changes. Applies cleanly against 2.6.11-rc1 and 2.6.11-rc1-bk6. Applies
  with offsets to 2.6.11-rc1-mm1

Changelog since V3
o inlined get_pageblock_type() and set_pageblock_type()
o set_pageblock_type() now takes a zone parameter to avoid a call to page_zone()
o When taking from the global pool, do not scan all the low-order lists

Changelog since V2
o Do not to interfere with the "min" decay
o Update the __GFP_BITS_SHIFT properly. Old value broke fsync and probably
  anything to do with asynchronous IO

Changelog since V1
o Update patch to 2.6.11-rc1
o Cleaned up bug where memory was wasted on a large bitmap
o Remove code that needed the binary buddy bitmaps
o Update flags to avoid colliding with __GFP_ZERO changes
o Extended fallback_count bean counters to show the fallback count for each
  allocation type
o In-code documentation

Version 1
o Initial release against 2.6.9

This patch is designed to reduce fragmentation in the standard buddy allocator
without impairing the performance of the allocator. High fragmentation in the
standard binary buddy allocator means that high-order allocations can rarely be
serviced which impacts HugeTLB allocations and the hot-removal of memory. This
patch works by dividing allocations into two different types of allocations;

EasyReclaimable - These are userspace pages that are easily reclaimable. This
	flag is set when it is known that the pages will be trivially reclaimed
	by writing the page out to swap or syncing with backing storage

KernelNonReclaimable - These are pages that are allocated by the kernel that
	are not trivially reclaimed. For example, the memory allocated for a
	loaded module would be in this category. By default, allocations are
	considered to be of this type

Instead of having one MAX_ORDER-sized array of free lists in struct free_area,
there are two, one for each type of allocation. Once a 2^MAX_ORDER block of
pages is split for a type of allocation, it is added to the free-lists for
that type, in effect reserving it. Hence, over time, pages of the different
types can be clustered together. When a page is allocated, the page-flags
are updated with a value indicating it's type of page so that it is placed
on the correct list on free.

When the preferred freelists are expired, the largest possible block is taken
from the alternative list. Buddies that are split from that large block are
placed on the preferred allocation-type freelists to mitigate fragmentation.

Three benchmark results are included all based on a 2.6.16-rc1-mm1 kernel
compiled with gcc 3.4. These benchmarks were run in the order you see them
*without* rebooting. This means that when the final highorder stress test,
the system is already running with any fragmentation introduced by other
benchmarks.

The first test called bench-kbuild.sh times a kernel build. Time is in seconds

                               2.6.16-rc1-mm1-clean  2.6.16-rc1-mm1-mbuddy-v22
Time taken to extract kernel:                    14                         15
Time taken to build kernel:                     741                        741

The second is the output of portions of AIM9 for the vanilla
allocator and the list-based anti-fragmentation one;

(Tests run with bench-aim9.sh from VMRegress 0.20)
                 2.6.16-rc1-mm1-clean  2.6.16-rc1-mm1-mbuddy-v22
 1 creat-clo                 12273.11                   12239.80     -33.31 -0.27% File Creations and Closes/second
 2 page_test                131762.75                  134311.90    2549.15 1.93% System Allocations & Pages/second
 3 brk_test                 586206.90                  597167.14   10960.24 1.87% System Memory Allocations/second
 4 jmp_test                4375520.75                 4373004.50   -2516.25 -0.06% Non-local gotos/second
 5 signal_test               79436.76                   77307.56   -2129.20 -2.68% Signal Traps/second
 6 exec_test                    62.90                      62.93       0.03 0.05% Program Loads/second
 7 fork_test                  1211.92                    1218.13       6.21 0.51% Task Creations/second
 8 link_test                  4332.30                    4324.56      -7.74 -0.18% Link/Unlink Pairs/second

The last test is to show that the allocator can satisfy more high-order
allocations, especially under load, than the standard allocator. The test
performs the following;

1. Start updatedb running in the background
2. Load kernel modules that tries to allocate high-order blocks on demand
3. Clean a kernel tree
4. Make 6 copies of the tree. As each copy finishes, a compile starts at -j1
5. Start compiling the primary tree
6. Sleep 1 minute while the 7 trees are being compiled
7. Use the kernel module to attempt 160 times to allocate a 2^10 block of pages
    - note, it only attempts 275 times, no matter how often it succeeds
    - An allocation is attempted every 1/10th of a second
    - Performance will get badly shot as it forces considerable amounts of
      pageout
8. At rest, dd a file from /dev/zero that is the size of physical memory, cat
   it and delete it again to flush as much of buffer cache as possible. Then
   try and allocate as many pages as possible

The result of the allocations under load;

HighAlloc Under Load Test Results Pass 1
                           2.6.16-rc1-mm1-clean  2.6.16-rc1-mm1-mbuddy-v22 
Order                                        10                         10 
Allocation type                         HighMem                    HighMem 
Attempted allocations                       275                        275 
Success allocs                               60                         86 
Failed allocs                               215                        189 
DMA zone allocs                               1                          1 
Normal zone allocs                            5                          0 
HighMem zone allocs                          54                         85 
EasyRclm zone allocs                          0                          0 
% Success                                    21                         31 

HighAlloc Under Load Test Results Pass 2
                           2.6.16-rc1-mm1-clean  2.6.16-rc1-mm1-mbuddy-v22 
Order                                        10                         10 
Allocation type                         HighMem                    HighMem 
Attempted allocations                       275                        275 
Success allocs                              101                        103 
Failed allocs                               174                        172 
DMA zone allocs                               1                          1 
Normal zone allocs                            5                          0 
HighMem zone allocs                          95                        102 
EasyRclm zone allocs                          0                          0 
% Success                                    36                         37 

HighAlloc Test Results while Rested
                           2.6.16-rc1-mm1-clean  2.6.16-rc1-mm1-mbuddy-v22 
Order                                        10                         10 
Allocation type                         HighMem                    HighMem 
Attempted allocations                       275                        275 
Success allocs                              141                        242 
Failed allocs                               134                         33 
DMA zone allocs                               1                          1 
Normal zone allocs                           16                         83 
HighMem zone allocs                         124                        158 
EasyRclm zone allocs                          0                          0 
% Success                                    51                         88 

Under load, it appears there is not a large differences in the percentage of
success but list-based anti-fragmentation was able to allocate 31 more HugeTLB
pages which is a significant improvement. The "Pass 2" results look similar
until you take into account that the -clean kernel is configured with PTEs
allocated from low memory. When CONFIG_HIGHPTE is set, it's success rate is
lower. Higher success rates with list-based depend heavily on the decisions
made by LRU. 

A second indicator of how well fragmentation is addressed is testing again
after the load decreases.  After the tests completed, the standard allocator
was able to allocate 141 order-10 pages and list-based anti-fragmentation
allocated 242, a massive improvement.

The results show that the modified allocator has comparable speed, but
is far less fragmented and in a better position to satisfy high-order
allocations. Unlike zone-based, it requires no configuration at boot time
to achieve the results.
-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

^ permalink raw reply	[flat|nested] 11+ messages in thread