tldr: Benchmarkers, only test patches 1-35.
git tree: git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux-balancenuma.git mm-balancenuma-v3r27
This is a large drop and is a bit more rushed than I'd like but delaying
it was not an option. This can be roughly considered to be in four major stages
1. Basic foundation, very similar to what was in V1
2. Full PMD fault handling, rate limiting of migration, two-stage migration filter.
This will migrate pages on a PTE or PMD level using just the current referencing
CPU as a placement hint
3. TLB flush optimisations
4. CPU follows memory algorithm. Very broadly speaking the intention is that based
on fault statistics a home node is identified and the process tries to remain
on the home node. It's crude and a much more complete implementation is needed.
Very broadly speaking the most urgent TODOs that spring to mind are
1. Move change_prot_numa to be based on change_protection
2. Native THP migration
3. Mitigate TLB flush operations in try_to_unmap_one called from migration path
4. Tunable to enable/disable from command-line and at runtime. It should be completely
disabled if the machine does not support NUMA.
5. Better load balancer integration (current is based on an old version of schednuma)
6. Fix/replace CPU follows algorithm. Current one is a broken port from autonuma, it's
very expensive and migrations are excessive. Either autonuma, schednuma or something
else needs to be rebased on top of this properly. The broken implementation gives
an indication where all the different parts should be plumbed in.
7. Depending on what happens with 6, fold page struct additions into page->flags
8. Revisit MPOL_NOOP and MPOL_MF_LAZY
9. Other architecture support or at least validation that it could be made work. I'm
half-hoping that the PPC64 people are watching because they tend to be interested
in this type of thing.
10. A review of all the conditionally compiled stuff. More of it could be compiled
out if !CONFIG_NUMA or !CONFIG_BALANCE_NUMA.
In terms of benchmarking only patches 1-35 should be considered. Patches
36-43 implement a placement policy that I know is not working as planned at
the moment. Note that all my own benchmarking did *not* include patch 16
"mm: mempolicy: Hide MPOL_NOOP and MPOL_MF_LAZY from userspace for now"
but that should not make a difference.
I'm leaving the RFC in place because patches 36-43 are incomplete.
Changelog since V2
o Do not allocate from home node
o Mostly remove pmd_numa handling for regular pmds
o HOME policy will allocate from and migrate towards local node
o Load balancer is more aggressive about moving tasks towards home node
o Renames to sync up more with -tip version
o Move pte handlers to generic code
o Scanning rate starts at 100ms, system CPU usage expected to increase
o Handle migration of PMD hinting faults
o Rate limit migration on a per-node basis
o Alter how the rate of PTE scanning is adapted
o Rate limit setting of pte_numa if node is congested
o Only flush local TLB is unmapping a pte_numa page
o Only consider one CPU in cpu follow algorithm
Changelog since V1
o Account for faults on the correct node after migration
o Do not account for THP splits as faults.
o Account THP faults on the node they occurred
o Ensure preferred_node_policy is initialised before use
o Mitigate double faults
o Add home-node logic
o Add some tlb-flush mitigation patches
o Add variation of CPU follows memory algorithm
o Add last_nid and use it as a two-stage filter before migrating pages
o Restart the PTE scanner when it reaches the end of the address space
o Lots of stuff I did not note properly
There are currently two competing approaches to implement support for
automatically migrating pages to optimise NUMA locality. Performance results
are available for both but review highlighted different problems in both.
They are not compatible with each other even though some fundamental
mechanics should have been the same. This series addresses part of the
integration and sharing problem by implementing a foundation that either
the policy for schednuma or autonuma can be rebased on.
The initial policy it implements is a very basic greedy policy called
"Migrate On Reference Of pte_numa Node (MORON)" and is later replaced by
a variation of the home-node policy and renamed. I expect to build upon
this revised policy and rename it to something more sensible that reflects
what it means.
In terms of building on top of the foundation the ideal would be that
patches affect one of the following areas although obviously that will
not always be possible
1. The PTE update helper functions
2. The PTE scanning machinary driven from task_numa_tick
3. Task and process fault accounting and how that information is used
to determine if a page is misplaced
4. Fault handling, migrating the page if misplaced, what information is
provided to the placement policy
5. Scheduler and load balancing
Patches 1-3 move some vmstat counters so that migrated pages get accounted
for. In the past the primary user of migration was compaction but
if pages are to migrate for NUMA optimisation then the counters
need to be generally useful.
Patch 4 defines an arch-specific PTE bit called _PAGE_NUMA that is used
to trigger faults later in the series. A placement policy is expected
to use these faults to determine if a page should migrate. On x86,
the bit is the same as _PAGE_PROTNONE but other architectures
may differ.
Patch 5-8 defines pte_numa, pmd_numa, pte_mknuma, pte_mknonuma and
friends. It implements them for x86, handles GUP and preserves
the _PAGE_NUMA bit across THP splits.
Patch 9 creates the fault handler for p[te|md]_numa PTEs and just clears
them again.
Patch 10 adds a MPOL_LOCAL policy so applications can explicitly request the
historical behaviour.
Patch 11 is premature but adds a MPOL_NOOP policy that can be used in
conjunction with the LAZY flags introduced later in the series.
Patch 12 adds migrate_misplaced_page which is responsible for migrating
a page to a new location.
Patch 13 migrates the page on fault if mpol_misplaced() says to do so.
Patch 14 updates the page fault handlers. Transparent huge pages are split.
Pages pointed to by PTEs are migrated. Pages pointed to by PMDs
are not properly handed until later in the series.
Patch 15 adds a MPOL_MF_LAZY mempolicy that an interested application can use.
On the next reference the memory should be migrated to the node that
references the memory.
Patch 16 notes that the MPOL_MF_LAZY and MPOL_NOOP flags have not been properly
reviewed and there are no manual pages. They are removed for now and
need to be revisited.
Patch 17 adds an arch flag for supporting balance numa
Patch 18 sets pte_numa within the context of the scheduler.
Patch 19 tries to avoid double faulting after migrating a page
Patches 20-22 note that the marking of pte_numa has a number of disadvantages and
instead incrementally updates a limited range of the address space
each tick.
Patch 23 adds some vmstats that can be used to approximate the cost of the
scheduling policy in a more fine-grained fashion than looking at
the system CPU usage.
Patch 24 implements the MORON policy. This is roughly where V1 of the series was.
Patch 25 properly handles the migration of pages faulted when handling a pmd
numa hinting fault. This could be improved as it's a bit tangled
to follow.
Patch 26 will only mark a PMD pmd_numa if many of the pages underneath are on
the same node.
Patches 27-29 rate-limit the number of pages being migrated and marked as pte_numa
Patch 30 slowly decreases the pte_numa update scanning rate
Patch 31-32 introduces last_nid and uses it to build a two-stage filter
that delays when a page gets migrated to avoid a situation where
a task running temporarily off its home node forces a migration.
Patches 33-35 brings in some TLB flush reduction patches. It was pointed
out that try_to_unmap_one still incurs a TLB flush and this is true.
An initial patch to cover this looked promising but was suspected
of a stability issue. It was likely triggered by another corruption
bug that has since been fixed and needs to be revisited.
Patches 36-39 introduces the concept of a home-node that the scheduler tries
to keep processes on. It's advisory only and not particularly strict.
There may be a problem with this whereby the load balancer is not
pushing processes back to their home node because there are no
idle CPUs available. It might need to be more aggressive about
swapping two tasks that are both running off their home node.
Patch 40 implements a CPU follow memory policy that is roughly based on what
was in autonuma. It builds statistics on faults on a per-task and
per-mm basis and decides if a tasks home node should be updated
on that basis. It is basically broken at the moment, is far too
heavy and results in bouncing but it serves as an illustration.
It needs to be reworked significantly or reimplemented.
Patch 41 makes patch 40 slightly less expensive but still way too heavy
Patch 42 adapts the pte_numa scanning rates based on the placement policy.
This also needs to be redone as it was found while writing this
changelog that the system CPU cost of reducing the scanning rate
is SEVERE. I kept the patch because it serves as a reminder that
we should do something like this.
Some notes.
This still is missing a mechanism for disabling from the command-line.
Documentation is sorely missing at this point.
In the past I noticed from profiles that mutex_spin_on_owner()
is very high in the last. I do not have recent profiles but will run
something over the weekend. The old observation was that on autonumabench
NUMA01_THREADLOCAL, the patches spend more time spinning in there and more
time in intel_idle implying that other users are waiting for the pte_numa
updates to complete. In the autonumabenchmark cases, the other contender
could be khugepaged. In the specjbb case there is also a lot of spinning
and it could be due to the JVM calling mprotect(). One way or the other,
it needs to be pinned down if the pte_numa updates are the problem and
if so how we might work around the requirement to hold mmap_sem while the
pte_numa update takes place.
Now the usual round of benchmarking! 7 kernels were considered, all based
on 3.7-rc4.
schednuma-v2r3 tip/sched/core + latest patches from Peter and Ingo
autonuma-v28fast rebased autonuma-v28fast branch from Andrea
stats-v2r34 Patches 1-3 of this series
moron-v3r27 Patches 1-24. MORON policy (similar to v1 of series)
twostage-v3r27 Patches 1-32. PMD handling, rate limiting, two-stage filter
lessflush-v3r27 Patches 1-35. TLB flush fixes on top
cpuone-v3r27 Patches 1-42. CPU follows algorithm
adaptscan-v3r27 Patches 1-43. Adaptive scanning
AUTONUMA BENCH
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34 rc4-schednuma-v2r3 rc4-autonuma-v28fast rc4-moron-v3r27 rc4-twostage-v3r27 rc4-lessflush-v3r27 rc4-cpuone-v3r27 rc4-adaptscan-v3r27
User NUMA01 67351.66 ( 0.00%) 47146.57 ( 30.00%) 30273.64 ( 55.05%) 23514.02 ( 65.09%) 62299.74 ( 7.50%) 66947.87 ( 0.60%) 55683.74 ( 17.32%) 40591.96 ( 39.73%)
User NUMA01_THEADLOCAL 54788.28 ( 0.00%) 17198.99 ( 68.61%) 17039.73 ( 68.90%) 20074.86 ( 63.36%) 22192.46 ( 59.49%) 21008.74 ( 61.65%) 18174.40 ( 66.83%) 17027.78 ( 68.92%)
User NUMA02 7179.87 ( 0.00%) 2096.07 ( 70.81%) 2099.85 ( 70.75%) 2902.95 ( 59.57%) 2140.49 ( 70.19%) 2208.52 ( 69.24%) 1125.91 ( 84.32%) 1329.20 ( 81.49%)
User NUMA02_SMT 3028.11 ( 0.00%) 998.22 ( 67.03%) 1052.97 ( 65.23%) 1051.16 ( 65.29%) 1053.06 ( 65.22%) 969.17 ( 67.99%) 778.44 ( 74.29%) 936.55 ( 69.07%)
System NUMA01 45.68 ( 0.00%) 3531.04 (-7629.95%) 423.91 (-828.00%) 723.05 (-1482.86%) 1548.99 (-3290.96%) 1903.18 (-4066.33%) 3762.31 (-8136.23%) 9143.26 (-19915.89%)
System NUMA01_THEADLOCAL 40.92 ( 0.00%) 926.72 (-2164.71%) 188.15 (-359.80%) 460.77 (-1026.03%) 685.06 (-1574.14%) 586.56 (-1333.43%) 1317.25 (-3119.09%) 4091.30 (-9898.29%)
System NUMA02 1.72 ( 0.00%) 23.64 (-1274.42%) 27.37 (-1491.28%) 33.15 (-1827.33%) 70.41 (-3993.60%) 72.02 (-4087.21%) 156.47 (-8997.09%) 158.89 (-9137.79%)
System NUMA02_SMT 0.92 ( 0.00%) 8.18 (-789.13%) 18.43 (-1903.26%) 22.31 (-2325.00%) 41.63 (-4425.00%) 38.06 (-4036.96%) 101.56 (-10939.13%) 65.32 (-7000.00%)
Elapsed NUMA01 1514.61 ( 0.00%) 1122.78 ( 25.87%) 722.66 ( 52.29%) 534.56 ( 64.71%) 1419.97 ( 6.25%) 1532.43 ( -1.18%) 1339.58 ( 11.56%) 1242.21 ( 17.98%)
Elapsed NUMA01_THEADLOCAL 1264.08 ( 0.00%) 393.79 ( 68.85%) 391.48 ( 69.03%) 471.07 ( 62.73%) 508.68 ( 59.76%) 487.97 ( 61.40%) 460.43 ( 63.58%) 531.53 ( 57.95%)
Elapsed NUMA02 181.88 ( 0.00%) 49.44 ( 72.82%) 61.55 ( 66.16%) 77.55 ( 57.36%) 60.96 ( 66.48%) 60.10 ( 66.96%) 56.96 ( 68.68%) 57.11 ( 68.60%)
Elapsed NUMA02_SMT 168.41 ( 0.00%) 47.49 ( 71.80%) 54.72 ( 67.51%) 66.98 ( 60.23%) 57.56 ( 65.82%) 54.06 ( 67.90%) 58.04 ( 65.54%) 53.99 ( 67.94%)
CPU NUMA01 4449.00 ( 0.00%) 4513.00 ( -1.44%) 4247.00 ( 4.54%) 4534.00 ( -1.91%) 4496.00 ( -1.06%) 4492.00 ( -0.97%) 4437.00 ( 0.27%) 4003.00 ( 10.02%)
CPU NUMA01_THEADLOCAL 4337.00 ( 0.00%) 4602.00 ( -6.11%) 4400.00 ( -1.45%) 4359.00 ( -0.51%) 4497.00 ( -3.69%) 4425.00 ( -2.03%) 4233.00 ( 2.40%) 3973.00 ( 8.39%)
CPU NUMA02 3948.00 ( 0.00%) 4287.00 ( -8.59%) 3455.00 ( 12.49%) 3785.00 ( 4.13%) 3626.00 ( 8.16%) 3794.00 ( 3.90%) 2251.00 ( 42.98%) 2605.00 ( 34.02%)
CPU NUMA02_SMT 1798.00 ( 0.00%) 2118.00 (-17.80%) 1957.00 ( -8.84%) 1602.00 ( 10.90%) 1901.00 ( -5.73%) 1862.00 ( -3.56%) 1516.00 ( 15.68%) 1855.00 ( -3.17%)
For NUMA01 moron-v3r27 does well but largely because it places things well
initially and then gets out of the way. The later patches in the series
do not cope as well. NUMA01 is an adverse workload and needs to be handled
better. The System CPU usage is high reflecting the migration it is doing
and while it's lower than schednuma's, it's still far too high.
In general, the System CPU usage is too high for everyone. Note that the
cpu follows algorithm puts it sky sky and the adaptive scanning makes it
worse. This needs addressing. A very large portion of this sytem CPU cost
is to due to TLB flushes during migration when handling pte_numa faults.
In terms of Elapsed time things are not too bad. For NUMA01_THEADLOCAL,
NUMA02 and NUMA02_SMT, lessflush-v3r27 (the main series that I think should
be benchmarked) shows reasonable improvements. It's not as good as schednuma
and autonuma in general but it is still respectable and there is no proper
placement policy after all.
MMTests Statistics: duration
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
stats-v2r34 schednuma-v2r3 autonuma moron-v3r27 twostage-v3r27 lessflush-v3 cpuone-v3 adaptscan-v3r27
User 132355.28 67445.10 50473.41 47550.20 87692.19 91141.69 75769.47 59892.77
System 89.90 4490.17 658.51 1239.92 2346.70 2600.48 5338.27 13459.43
Elapsed 3138.98 1621.73 1240.09 1159.42 2055.92 2144.42 1924.10 1893.29
Bit mushed up but the main take-away here is the System CPU
cost. wwostage-v3r27 and lessflush-v3r27 is very high and the placement
policy and adaptive scan make it a lot worse. autonumas looks really low
but this could be due to the fact it does a lot of work in kernel threads
where the cost is not as obvious.
MMTests Statistics: vmstat
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34rc4-schednuma-v2r3rc4-autonuma-v28fastrc4-moron-v3r27rc4-twostage-v3r27rc4-lessflush-v3r27rc4-cpuone-v3r27rc4-adaptscan-v3r27
Page Ins 40180 36944 41824 43420 43432 43168 43424 43148
Page Outs 29548 16996 13352 12864 20684 21628 17964 18984
Swap Ins 0 0 0 0 0 0 0 0
Swap Outs 0 0 0 0 0 0 0 0
Direct pages scanned 0 0 0 0 0 0 0 0
Kswapd pages scanned 0 0 0 0 0 0 0 0
Kswapd pages reclaimed 0 0 0 0 0 0 0 0
Direct pages reclaimed 0 0 0 0 0 0 0 0
Kswapd efficiency 100% 100% 100% 100% 100% 100% 100% 100%
Kswapd velocity 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Direct efficiency 100% 100% 100% 100% 100% 100% 100% 100%
Direct velocity 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Percentage direct scans 0% 0% 0% 0% 0% 0% 0% 0%
Page writes by reclaim 0 0 0 0 0 0 0 0
Page writes file 0 0 0 0 0 0 0 0
Page writes anon 0 0 0 0 0 0 0 0
Page reclaim immediate 0 0 0 0 0 0 0 0
Page rescued immediate 0 0 0 0 0 0 0 0
Slabs scanned 0 0 0 0 0 0 0 0
Direct inode steals 0 0 0 0 0 0 0 0
Kswapd inode steals 0 0 0 0 0 0 0 0
Kswapd skipped wait 0 0 0 0 0 0 0 0
THP fault alloc 16688 12225 19232 17117 17828 17273 18272 18695
THP collapse alloc 8 1 9743 484 918 1034 1097 1095
THP splits 3 0 10654 7568 7453 7679 8051 8134
THP fault fallback 0 0 0 0 0 0 0 0
THP collapse fail 0 0 0 0 0 0 0 0
Compaction stalls 0 0 0 0 0 0 0 0
Compaction success 0 0 0 0 0 0 0 0
Compaction failures 0 0 0 0 0 0 0 0
Page migrate success 0 0 0 3372219 9296248 9122453 19833353 42345720
Page migrate failure 0 0 0 0 0 0 0 0
Compaction pages isolated 0 0 0 0 0 0 0 0
Compaction migrate scanned 0 0 0 0 0 0 0 0
Compaction free scanned 0 0 0 0 0 0 0 0
Compaction cost 0 0 0 3500 9649 9469 20587 43954
NUMA PTE updates 0 0 0 571770975 101066122 104353617 411858897 1471434472
NUMA hint faults 0 0 0 573525212 103538510 106870459 415594465 1481522643
NUMA hint local faults 0 0 0 149965397 49345272 51268932 202046567 555527366
NUMA pages migrated 0 0 0 3372219 9296248 9122453 19833353 42345720
AutoNUMA cost 0 0 0 2871692 518576 535256 2081232 7418717
schednuma and autonuma do not have the stats so we cannot compare the
notional costs except to note that schednuma has no THP splits as it
supports native THP migration.
For balancenuma, the main thing to spot is that there are a LOT of pte
updates and migrations. Superficially this indicates that the workload is
not converging properly and reducing the scanning rate when it does. This
is where a proper placement policy, scheduling decisions and scan rate
adaption should come in to play.
SPECJBB BOPS
Cutting this one a bit short again to save pace
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34 rc4-schednuma-v2r3 rc4-autonuma-v28fast rc4-moron-v3r27 rc4-twostage-v3r27 rc4-lessflush-v3r27 rc4-cpuone-v3r27 rc4-adaptscan-v3r27
Mean 1 25034.25 ( 0.00%) 20598.50 (-17.72%) 25192.25 ( 0.63%) 25168.75 ( 0.54%) 25525.75 ( 1.96%) 25146.25 ( 0.45%) 24270.25 ( -3.05%) 24703.75 ( -1.32%)
Mean 2 53176.00 ( 0.00%) 43906.50 (-17.43%) 55508.25 ( 4.39%) 52909.50 ( -0.50%) 49688.25 ( -6.56%) 50920.75 ( -4.24%) 51530.25 ( -3.09%) 47131.25 (-11.37%)
Mean 3 77350.50 ( 0.00%) 60342.75 (-21.99%) 82122.50 ( 6.17%) 76557.00 ( -1.03%) 75532.25 ( -2.35%) 73726.75 ( -4.68%) 74520.00 ( -3.66%) 63232.25 (-18.25%)
Mean 4 99919.50 ( 0.00%) 80781.75 (-19.15%) 107233.25 ( 7.32%) 98943.50 ( -0.98%) 97165.00 ( -2.76%) 96004.75 ( -3.92%) 95784.00 ( -4.14%) 67122.00 (-32.82%)
Mean 5 119797.00 ( 0.00%) 97870.00 (-18.30%) 131016.00 ( 9.37%) 118532.25 ( -1.06%) 117484.50 ( -1.93%) 116902.50 ( -2.42%) 116510.25 ( -2.74%) 69520.00 (-41.97%)
Mean 6 135858.00 ( 0.00%) 123912.50 ( -8.79%) 152444.75 ( 12.21%) 133761.75 ( -1.54%) 133725.25 ( -1.57%) 134317.50 ( -1.13%) 132445.75 ( -2.51%) 42056.75 (-69.04%)
Mean 7 136074.00 ( 0.00%) 126574.25 ( -6.98%) 157372.75 ( 15.65%) 133870.75 ( -1.62%) 135822.50 ( -0.18%) 137850.25 ( 1.31%) 135727.75 ( -0.25%) 19630.75 (-85.57%)
Mean 8 132426.25 ( 0.00%) 121766.00 ( -8.05%) 161655.25 ( 22.07%) 131605.50 ( -0.62%) 136697.25 ( 3.23%) 135818.00 ( 2.56%) 135559.00 ( 2.37%) 27267.75 (-79.41%)
Mean 9 129432.75 ( 0.00%) 114224.25 (-11.75%) 160530.50 ( 24.03%) 130498.50 ( 0.82%) 134121.00 ( 3.62%) 133703.25 ( 3.30%) 134048.75 ( 3.57%) 18777.00 (-85.49%)
Mean 10 118399.75 ( 0.00%) 109040.50 ( -7.90%) 158692.00 ( 34.03%) 125355.50 ( 5.87%) 131581.75 ( 11.13%) 129295.50 ( 9.20%) 130685.25 ( 10.38%) 4565.00 (-96.14%)
Mean 11 119604.00 ( 0.00%) 105566.50 (-11.74%) 154462.00 ( 29.14%) 126155.75 ( 5.48%) 127086.00 ( 6.26%) 125634.25 ( 5.04%) 125426.00 ( 4.87%) 4811.25 (-95.98%)
Mean 12 112742.25 ( 0.00%) 101728.75 ( -9.77%) 149546.00 ( 32.64%) 111419.00 ( -1.17%) 118136.75 ( 4.78%) 118625.75 ( 5.22%) 120271.25 ( 6.68%) 5029.25 (-95.54%)
Mean 13 109480.75 ( 0.00%) 103737.50 ( -5.25%) 144929.25 ( 32.38%) 109388.25 ( -0.08%) 114351.25 ( 4.45%) 115157.00 ( 5.18%) 117752.00 ( 7.55%) 3314.25 (-96.97%)
Mean 14 109724.00 ( 0.00%) 103516.00 ( -5.66%) 143804.50 ( 31.06%) 108902.25 ( -0.75%) 115422.75 ( 5.19%) 114151.75 ( 4.04%) 115994.00 ( 5.71%) 5312.50 (-95.16%)
Mean 15 109111.75 ( 0.00%) 100817.00 ( -7.60%) 141878.00 ( 30.03%) 108213.25 ( -0.82%) 115640.00 ( 5.98%) 112870.00 ( 3.44%) 117017.00 ( 7.25%) 5298.75 (-95.14%)
Mean 16 105385.75 ( 0.00%) 99327.25 ( -5.75%) 140156.75 ( 32.99%) 105159.75 ( -0.21%) 113128.25 ( 7.35%) 111836.50 ( 6.12%) 115170.00 ( 9.28%) 4091.75 (-96.12%)
Mean 17 101903.50 ( 0.00%) 96464.50 ( -5.34%) 138402.00 ( 35.82%) 104582.75 ( 2.63%) 112576.00 ( 10.47%) 112967.50 ( 10.86%) 113390.75 ( 11.27%) 5601.25 (-94.50%)
Mean 18 103632.50 ( 0.00%) 95632.50 ( -7.72%) 137781.50 ( 32.95%) 103168.00 ( -0.45%) 110462.00 ( 6.59%) 113622.75 ( 9.64%) 113209.00 ( 9.24%) 6216.75 (-94.00%)
Stddev 1 1195.76 ( 0.00%) 358.07 ( 70.06%) 861.97 ( 27.91%) 1108.27 ( 7.32%) 704.35 ( 41.10%) 738.31 ( 38.26%) 370.96 ( 68.98%) 858.14 ( 28.23%)
Stddev 2 883.39 ( 0.00%) 1203.29 (-36.21%) 855.08 ( 3.20%) 320.44 ( 63.73%) 1190.25 (-34.74%) 918.86 ( -4.02%) 720.67 ( 18.42%) 1831.94 (-107.38%)
Stddev 3 997.25 ( 0.00%) 3755.67 (-276.60%) 545.50 ( 45.30%) 971.40 ( 2.59%) 1444.69 (-44.87%) 1507.91 (-51.21%) 1227.37 (-23.08%) 4043.37 (-305.45%)
Stddev 4 1115.16 ( 0.00%) 6390.65 (-473.07%) 1183.49 ( -6.13%) 679.74 ( 39.05%) 1320.08 (-18.38%) 897.64 ( 19.51%) 1525.30 (-36.78%) 8637.27 (-674.53%)
Stddev 5 1367.09 ( 0.00%) 9710.70 (-610.32%) 1022.09 ( 25.24%) 944.31 ( 30.93%) 1003.82 ( 26.57%) 824.03 ( 39.72%) 1128.73 ( 17.44%) 13504.42 (-887.82%)
Stddev 6 1125.22 ( 0.00%) 1097.83 ( 2.43%) 1013.52 ( 9.93%) 1170.85 ( -4.06%) 1971.57 (-75.22%) 1042.93 ( 7.31%) 2416.06 (-114.72%) 9214.24 (-718.89%)
Stddev 7 3211.72 ( 0.00%) 1533.62 ( 52.25%) 512.61 ( 84.04%) 4186.42 (-30.35%) 5832.10 (-81.59%) 4264.34 (-32.77%) 2886.05 ( 10.14%) 2628.35 ( 18.16%)
Stddev 8 4194.96 ( 0.00%) 1518.26 ( 63.81%) 493.64 ( 88.23%) 2203.56 ( 47.47%) 1961.15 ( 53.25%) 2913.42 ( 30.55%) 3445.70 ( 17.86%) 13053.31 (-211.17%)
Stddev 9 6175.10 ( 0.00%) 2648.75 ( 57.11%) 2109.83 ( 65.83%) 2732.83 ( 55.74%) 2205.91 ( 64.28%) 3808.45 ( 38.33%) 3246.22 ( 47.43%) 5511.26 ( 10.75%)
Stddev 10 4754.87 ( 0.00%) 1941.47 ( 59.17%) 2948.98 ( 37.98%) 1533.87 ( 67.74%) 2395.65 ( 49.62%) 3207.51 ( 32.54%) 3564.21 ( 25.04%) 783.51 ( 83.52%)
Stddev 11 2706.18 ( 0.00%) 1247.95 ( 53.89%) 5907.16 (-118.28%) 3030.54 (-11.99%) 2989.54 (-10.47%) 2983.44 (-10.25%) 3156.67 (-16.65%) 939.68 ( 65.28%)
Stddev 12 3607.76 ( 0.00%) 663.63 ( 81.61%) 9063.28 (-151.22%) 3191.77 ( 11.53%) 2849.20 ( 21.03%) 1810.51 ( 49.82%) 3422.89 ( 5.12%) 305.09 ( 91.54%)
Stddev 13 2771.67 ( 0.00%) 1447.87 ( 47.76%) 8716.51 (-214.49%) 3516.13 (-26.86%) 1425.69 ( 48.56%) 2564.87 ( 7.46%) 1667.33 ( 39.84%) 118.01 ( 95.74%)
Stddev 14 2522.18 ( 0.00%) 1510.28 ( 40.12%) 9286.98 (-268.21%) 3144.22 (-24.66%) 1866.90 ( 25.98%) 784.45 ( 68.90%) 369.15 ( 85.36%) 764.26 ( 69.70%)
Stddev 15 2711.16 ( 0.00%) 1719.54 ( 36.58%) 9895.88 (-265.01%) 2889.53 ( -6.58%) 1059.84 ( 60.91%) 2043.26 ( 24.64%) 1149.45 ( 57.60%) 297.90 ( 89.01%)
Stddev 16 2797.21 ( 0.00%) 983.63 ( 64.84%) 9302.92 (-232.58%) 2734.35 ( 2.25%) 817.51 ( 70.77%) 937.10 ( 66.50%) 1031.85 ( 63.11%) 223.38 ( 92.01%)
Stddev 17 4019.85 ( 0.00%) 1927.25 ( 52.06%) 9998.34 (-148.72%) 2567.94 ( 36.12%) 1301.02 ( 67.64%) 1803.98 ( 55.12%) 1683.85 ( 58.11%) 697.06 ( 82.66%)
Stddev 18 3332.20 ( 0.00%) 1401.68 ( 57.94%) 12056.08 (-261.80%) 2297.48 ( 31.05%) 1852.32 ( 44.41%) 675.02 ( 79.74%) 1190.98 ( 64.26%) 285.90 ( 91.42%)
TPut 1 100137.00 ( 0.00%) 82394.00 (-17.72%) 100769.00 ( 0.63%) 100675.00 ( 0.54%) 102103.00 ( 1.96%) 100585.00 ( 0.45%) 97081.00 ( -3.05%) 98815.00 ( -1.32%)
TPut 2 212704.00 ( 0.00%) 175626.00 (-17.43%) 222033.00 ( 4.39%) 211638.00 ( -0.50%) 198753.00 ( -6.56%) 203683.00 ( -4.24%) 206121.00 ( -3.09%) 188525.00 (-11.37%)
TPut 3 309402.00 ( 0.00%) 241371.00 (-21.99%) 328490.00 ( 6.17%) 306228.00 ( -1.03%) 302129.00 ( -2.35%) 294907.00 ( -4.68%) 298080.00 ( -3.66%) 252929.00 (-18.25%)
TPut 4 399678.00 ( 0.00%) 323127.00 (-19.15%) 428933.00 ( 7.32%) 395774.00 ( -0.98%) 388660.00 ( -2.76%) 384019.00 ( -3.92%) 383136.00 ( -4.14%) 268488.00 (-32.82%)
TPut 5 479188.00 ( 0.00%) 391480.00 (-18.30%) 524064.00 ( 9.37%) 474129.00 ( -1.06%) 469938.00 ( -1.93%) 467610.00 ( -2.42%) 466041.00 ( -2.74%) 278080.00 (-41.97%)
TPut 6 543432.00 ( 0.00%) 495650.00 ( -8.79%) 609779.00 ( 12.21%) 535047.00 ( -1.54%) 534901.00 ( -1.57%) 537270.00 ( -1.13%) 529783.00 ( -2.51%) 168227.00 (-69.04%)
TPut 7 544296.00 ( 0.00%) 506297.00 ( -6.98%) 629491.00 ( 15.65%) 535483.00 ( -1.62%) 543290.00 ( -0.18%) 551401.00 ( 1.31%) 542911.00 ( -0.25%) 78523.00 (-85.57%)
TPut 8 529705.00 ( 0.00%) 487064.00 ( -8.05%) 646621.00 ( 22.07%) 526422.00 ( -0.62%) 546789.00 ( 3.23%) 543272.00 ( 2.56%) 542236.00 ( 2.37%) 109071.00 (-79.41%)
TPut 9 517731.00 ( 0.00%) 456897.00 (-11.75%) 642122.00 ( 24.03%) 521994.00 ( 0.82%) 536484.00 ( 3.62%) 534813.00 ( 3.30%) 536195.00 ( 3.57%) 75108.00 (-85.49%)
TPut 10 473599.00 ( 0.00%) 436162.00 ( -7.90%) 634768.00 ( 34.03%) 501422.00 ( 5.87%) 526327.00 ( 11.13%) 517182.00 ( 9.20%) 522741.00 ( 10.38%) 18260.00 (-96.14%)
TPut 11 478416.00 ( 0.00%) 422266.00 (-11.74%) 617848.00 ( 29.14%) 504623.00 ( 5.48%) 508344.00 ( 6.26%) 502537.00 ( 5.04%) 501704.00 ( 4.87%) 19245.00 (-95.98%)
TPut 12 450969.00 ( 0.00%) 406915.00 ( -9.77%) 598184.00 ( 32.64%) 445676.00 ( -1.17%) 472547.00 ( 4.78%) 474503.00 ( 5.22%) 481085.00 ( 6.68%) 20117.00 (-95.54%)
TPut 13 437923.00 ( 0.00%) 414950.00 ( -5.25%) 579717.00 ( 32.38%) 437553.00 ( -0.08%) 457405.00 ( 4.45%) 460628.00 ( 5.18%) 471008.00 ( 7.55%) 13257.00 (-96.97%)
TPut 14 438896.00 ( 0.00%) 414064.00 ( -5.66%) 575218.00 ( 31.06%) 435609.00 ( -0.75%) 461691.00 ( 5.19%) 456607.00 ( 4.04%) 463976.00 ( 5.71%) 21250.00 (-95.16%)
TPut 15 436447.00 ( 0.00%) 403268.00 ( -7.60%) 567512.00 ( 30.03%) 432853.00 ( -0.82%) 462560.00 ( 5.98%) 451480.00 ( 3.44%) 468068.00 ( 7.25%) 21195.00 (-95.14%)
TPut 16 421543.00 ( 0.00%) 397309.00 ( -5.75%) 560627.00 ( 32.99%) 420639.00 ( -0.21%) 452513.00 ( 7.35%) 447346.00 ( 6.12%) 460680.00 ( 9.28%) 16367.00 (-96.12%)
TPut 17 407614.00 ( 0.00%) 385858.00 ( -5.34%) 553608.00 ( 35.82%) 418331.00 ( 2.63%) 450304.00 ( 10.47%) 451870.00 ( 10.86%) 453563.00 ( 11.27%) 22405.00 (-94.50%)
TPut 18 414530.00 ( 0.00%) 382530.00 ( -7.72%) 551126.00 ( 32.95%) 412672.00 ( -0.45%) 441848.00 ( 6.59%) 454491.00 ( 9.64%) 452836.00 ( 9.24%) 24867.00 (-94.00%)
One JVM runs per numa node. Mean is average ops/sec per JVM. Tput is
overall throughput of all nodes.
lessflush-v3r27 does reasonably well here. It's slower for smaller number of
warehouses and sees 3-10% performance gains for larger numbers of warehouses.
This is quite encouraging. Note that moron-v3r27 which is roughly similar
to v1 of this series is crap because of its brain-damaged handling of
PMD faults.
The cpu-follows policy does nothing useful here. If it's making better placement
decisions, it's losing all the gain.
The adaptive scan COMPLETELY wrecks everything. I was tempted to delete this patch
entirely and pretend it didn't exist but some sort of adaptive scan rate is required.
The patch at least acts as a "Don't Do What Donny Don't Did".
schednuma regressses badly here and it has to be established why as Ingo reports
the exact opposite. It has been discussed elsewhere but it could be down to the
kernel, the machine, the JVM configuration or which specjbb figures we are
actually reporting.
schednuma and lessflush-v3r27 are reasonably good in terms of variations
across JVMs and is generally more. autonuma has very variable performance between JVMs.
autonuma dominates here.
SPECJBB PEAKS
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34 rc4-schednuma-v2r3 rc4-autonuma-v28fast rc4-moron-v3r27 rc4-twostage-v3r27 rc4-lessflush-v3r27 rc4-cpuone-v3r27 rc4-adaptscan-v3r27
Expctd Warehouse 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%) 12.00 ( 0.00%)
Expctd Peak Bops 450969.00 ( 0.00%) 406915.00 ( -9.77%) 598184.00 ( 32.64%) 445676.00 ( -1.17%) 472547.00 ( 4.78%) 474503.00 ( 5.22%) 481085.00 ( 6.68%) 20117.00 (-95.54%)
Actual Warehouse 7.00 ( 0.00%) 7.00 ( 0.00%) 8.00 ( 14.29%) 7.00 ( 0.00%) 8.00 ( 14.29%) 7.00 ( 0.00%) 7.00 ( 0.00%) 5.00 (-28.57%)
Actual Peak Bops 544296.00 ( 0.00%) 506297.00 ( -6.98%) 646621.00 ( 18.80%) 535483.00 ( -1.62%) 546789.00 ( 0.46%) 551401.00 ( 1.31%) 542911.00 ( -0.25%) 278080.00 (-48.91%)
Other than autonuma, peak performance did not go well. balancenuma
sustains performance for greater numbers of warehouses but it's actual
peak performance is not improved. As before, adaptive scan killed everything.
MMTests Statistics: duration
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34rc4-schednuma-v2r3rc4-autonuma-v28fastrc4-moron-v3r27rc4-twostage-v3r27rc4-lessflush-v3r27rc4-cpuone-v3r27rc4-adaptscan-v3r27
User 101949.84 86817.79 101748.80 100943.56 99799.41 99896.98 99813.11 12790.74
System 66.05 13094.99 191.40 948.00 1948.39 1939.91 1995.15 40647.38
Elapsed 2456.35 2459.16 2451.96 2456.83 2462.20 2462.01 2462.97 2502.24
schednumas system CPU costs were high.
autonumas were low but again, the cost could be hidden.
balancenumas is relatively not too bad (other than adaptive scan which
kills the world) but it is still stupidly high. A proper placement policy
that reduced migrations would help a lot.
MMTests Statistics: vmstat
3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0 3.7.0
rc4-stats-v2r34rc4-schednuma-v2r3rc4-autonuma-v28fastrc4-moron-v3r27rc4-twostage-v3r27rc4-lessflush-v3r27rc4-cpuone-v3r27rc4-adaptscan-v3r27
Page Ins 34920 36128 37356 38264 38368 37952 38196 38236
Page Outs 32116 34000 31140 31604 31152 32872 31592 33280
Swap Ins 0 0 0 0 0 0 0 0
Swap Outs 0 0 0 0 0 0 0 0
Direct pages scanned 0 0 0 0 0 0 0 0
Kswapd pages scanned 0 0 0 0 0 0 0 0
Kswapd pages reclaimed 0 0 0 0 0 0 0 0
Direct pages reclaimed 0 0 0 0 0 0 0 0
Kswapd efficiency 100% 100% 100% 100% 100% 100% 100% 100%
Kswapd velocity 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Direct efficiency 100% 100% 100% 100% 100% 100% 100% 100%
Direct velocity 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Percentage direct scans 0% 0% 0% 0% 0% 0% 0% 0%
Page writes by reclaim 0 0 0 0 0 0 0 0
Page writes file 0 0 0 0 0 0 0 0
Page writes anon 0 0 0 0 0 0 0 0
Page reclaim immediate 0 0 0 0 0 0 0 0
Page rescued immediate 0 0 0 0 0 0 0 0
Slabs scanned 0 0 0 0 0 0 0 0
Direct inode steals 0 0 0 0 0 0 0 0
Kswapd inode steals 0 0 0 0 0 0 0 0
Kswapd skipped wait 0 0 0 0 0 0 0 0
THP fault alloc 1 1 1 2 1 2 1 2
THP collapse alloc 0 0 23 0 0 0 0 2
THP splits 0 0 7 0 3 1 7 5
THP fault fallback 0 0 0 0 0 0 0 0
THP collapse fail 0 0 0 0 0 0 0 0
Compaction stalls 0 0 0 0 0 0 0 0
Compaction success 0 0 0 0 0 0 0 0
Compaction failures 0 0 0 0 0 0 0 0
Page migrate success 0 0 0 890168 53347569 53708970 53869395 381749347
Page migrate failure 0 0 0 0 0 0 0 0
Compaction pages isolated 0 0 0 0 0 0 0 0
Compaction migrate scanned 0 0 0 0 0 0 0 0
Compaction free scanned 0 0 0 0 0 0 0 0
Compaction cost 0 0 0 923 55374 55749 55916 396255
NUMA PTE updates 0 0 0 2959462982 382203645 383516027 388145421 3175106653
NUMA hint faults 0 0 0 2958118854 381790747 382914344 387738802 3202932515
NUMA hint local faults 0 0 0 771705175 102887391 104071278 102032245 1038500179
NUMA pages migrated 0 0 0 890168 53347569 53708970 53869395 381749347
AutoNUMA cost 0 0 0 14811327 1912642 1918276 1942434 16044141
THP is not really a factor for this workload but one thing to note is the
migration rate for lessflush-v3r27. It works out at migrating 85MB/s on
average throughout the entire test. Again, a proper placement policy
should reduce this.
So in summary, patches 1-35 are not perfect and needs a proper placement
policy and scheduler smarts but out of the box it's not completely crap
either.
arch/sh/mm/Kconfig | 1 +
arch/x86/Kconfig | 1 +
arch/x86/include/asm/pgtable.h | 11 +-
arch/x86/include/asm/pgtable_types.h | 20 +
arch/x86/mm/pgtable.c | 8 +-
include/asm-generic/pgtable.h | 7 +
include/linux/huge_mm.h | 10 +
include/linux/init_task.h | 8 +
include/linux/mempolicy.h | 8 +
include/linux/migrate.h | 27 +-
include/linux/mm.h | 34 ++
include/linux/mm_types.h | 44 ++
include/linux/mmzone.h | 13 +
include/linux/sched.h | 52 +++
include/linux/vm_event_item.h | 12 +-
include/linux/vmstat.h | 8 +
include/trace/events/migrate.h | 51 +++
include/uapi/linux/mempolicy.h | 24 +-
init/Kconfig | 22 +
kernel/fork.c | 18 +
kernel/sched/core.c | 60 ++-
kernel/sched/debug.c | 3 +
kernel/sched/fair.c | 764 ++++++++++++++++++++++++++++++++--
kernel/sched/features.h | 25 ++
kernel/sched/sched.h | 36 ++
kernel/sysctl.c | 38 +-
mm/compaction.c | 15 +-
mm/huge_memory.c | 53 +++
mm/memory-failure.c | 3 +-
mm/memory.c | 198 ++++++++-
mm/memory_hotplug.c | 3 +-
mm/mempolicy.c | 381 +++++++++++++++--
mm/migrate.c | 178 +++++++-
mm/page_alloc.c | 10 +-
mm/pgtable-generic.c | 59 ++-
mm/vmstat.c | 16 +-
36 files changed, 2131 insertions(+), 90 deletions(-)
create mode 100644 include/trace/events/migrate.h
--
1.7.9.2
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>