Many workloads that allocate pages are not handling an interrupt at a
time. As allocation requests may be from IRQ context, it's necessary to
disable/enable IRQs for every page allocation. This cost is the bulk of
the free path but also a significant percentage of the allocation path.
This patch alters the locking and checks such that only irq-safe
allocation requests use the per-cpu allocator. All others acquire the
irq-safe zone->lock and allocate from the buddy allocator. It relies on
disabling preemption to safely access the per-cpu structures. It could
be slightly modified to avoid soft IRQs using it but it's not clear it's
worthwhile.
This modification may slow allocations from IRQ context slightly but the
main gain from the per-cpu allocator is that it scales better for
allocations from multiple contexts. There is an implicit assumption
that intensive allocations from IRQ contexts on multiple CPUs from a
single NUMA node are rare and that the fast majority of scaling issues
are encountered in !IRQ contexts such as page faulting. It's worth
noting that this patch is not required for a bulk page allocator but it
significantly reduces the overhead.
The following is results from a page allocator micro-benchmark. Only
order-0 is interesting as higher orders do not use the per-cpu allocator
4.10.0-rc2 4.10.0-rc2
vanilla irqsafe-v1r5
Amean alloc-odr0-1 287.15 ( 0.00%) 219.00 ( 23.73%)
Amean alloc-odr0-2 221.23 ( 0.00%) 183.23 ( 17.18%)
Amean alloc-odr0-4 187.00 ( 0.00%) 151.38 ( 19.05%)
Amean alloc-odr0-8 167.54 ( 0.00%) 132.77 ( 20.75%)
Amean alloc-odr0-16 156.00 ( 0.00%) 123.00 ( 21.15%)
Amean alloc-odr0-32 149.00 ( 0.00%) 118.31 ( 20.60%)
Amean alloc-odr0-64 138.77 ( 0.00%) 116.00 ( 16.41%)
Amean alloc-odr0-128 145.00 ( 0.00%) 118.00 ( 18.62%)
Amean alloc-odr0-256 136.15 ( 0.00%) 125.00 ( 8.19%)
Amean alloc-odr0-512 147.92 ( 0.00%) 121.77 ( 17.68%)
Amean alloc-odr0-1024 147.23 ( 0.00%) 126.15 ( 14.32%)
Amean alloc-odr0-2048 155.15 ( 0.00%) 129.92 ( 16.26%)
Amean alloc-odr0-4096 164.00 ( 0.00%) 136.77 ( 16.60%)
Amean alloc-odr0-8192 166.92 ( 0.00%) 138.08 ( 17.28%)
Amean alloc-odr0-16384 159.00 ( 0.00%) 138.00 ( 13.21%)
Amean free-odr0-1 165.00 ( 0.00%) 89.00 ( 46.06%)
Amean free-odr0-2 113.00 ( 0.00%) 63.00 ( 44.25%)
Amean free-odr0-4 99.00 ( 0.00%) 54.00 ( 45.45%)
Amean free-odr0-8 88.00 ( 0.00%) 47.38 ( 46.15%)
Amean free-odr0-16 83.00 ( 0.00%) 46.00 ( 44.58%)
Amean free-odr0-32 80.00 ( 0.00%) 44.38 ( 44.52%)
Amean free-odr0-64 72.62 ( 0.00%) 43.00 ( 40.78%)
Amean free-odr0-128 78.00 ( 0.00%) 42.00 ( 46.15%)
Amean free-odr0-256 80.46 ( 0.00%) 57.00 ( 29.16%)
Amean free-odr0-512 96.38 ( 0.00%) 64.69 ( 32.88%)
Amean free-odr0-1024 107.31 ( 0.00%) 72.54 ( 32.40%)
Amean free-odr0-2048 108.92 ( 0.00%) 78.08 ( 28.32%)
Amean free-odr0-4096 113.38 ( 0.00%) 82.23 ( 27.48%)
Amean free-odr0-8192 112.08 ( 0.00%) 82.85 ( 26.08%)
Amean free-odr0-16384 110.38 ( 0.00%) 81.92 ( 25.78%)
Amean total-odr0-1 452.15 ( 0.00%) 308.00 ( 31.88%)
Amean total-odr0-2 334.23 ( 0.00%) 246.23 ( 26.33%)
Amean total-odr0-4 286.00 ( 0.00%) 205.38 ( 28.19%)
Amean total-odr0-8 255.54 ( 0.00%) 180.15 ( 29.50%)
Amean total-odr0-16 239.00 ( 0.00%) 169.00 ( 29.29%)
Amean total-odr0-32 229.00 ( 0.00%) 162.69 ( 28.96%)
Amean total-odr0-64 211.38 ( 0.00%) 159.00 ( 24.78%)
Amean total-odr0-128 223.00 ( 0.00%) 160.00 ( 28.25%)
Amean total-odr0-256 216.62 ( 0.00%) 182.00 ( 15.98%)
Amean total-odr0-512 244.31 ( 0.00%) 186.46 ( 23.68%)
Amean total-odr0-1024 254.54 ( 0.00%) 198.69 ( 21.94%)
Amean total-odr0-2048 264.08 ( 0.00%) 208.00 ( 21.24%)
Amean total-odr0-4096 277.38 ( 0.00%) 219.00 ( 21.05%)
Amean total-odr0-8192 279.00 ( 0.00%) 220.92 ( 20.82%)
Amean total-odr0-16384 269.38 ( 0.00%) 219.92 ( 18.36%)
This is the alloc, free and total overhead of allocating order-0 pages
in batches of 1 page up to 16384 pages. Avoiding disabling/enabling
overhead massively reduces overhead. Alloc overhead is roughly reduced
by 14-20% in most cases. The free path is reduced by 26-46% and the
total reduction is significant.
Many users require zeroing of pages from the page allocator which is the
vast cost of allocation. Hence, the impact on a basic page faulting
benchmark is not that significant
4.10.0-rc2 4.10.0-rc2
vanilla irqsafe-v1r5
Hmean page_test 656632.98 ( 0.00%) 675536.13 ( 2.88%)
Hmean brk_test 3845502.67 ( 0.00%) 3867186.94 ( 0.56%)
Stddev page_test 10543.29 ( 0.00%) 4104.07 ( 61.07%)
Stddev brk_test 33472.36 ( 0.00%) 15538.39 ( 53.58%)
CoeffVar page_test 1.61 ( 0.00%) 0.61 ( 62.15%)
CoeffVar brk_test 0.87 ( 0.00%) 0.40 ( 53.84%)
Max page_test 666513.33 ( 0.00%) 678640.00 ( 1.82%)
Max brk_test 3882800.00 ( 0.00%) 3887008.66 ( 0.11%)
This is from aim9 and the most notable outcome is that fault variability
is reduced by the patch. The headline improvement is small as the
overall fault cost, zeroing, page table insertion etc dominate relative
to disabling/enabling IRQs in the per-cpu allocator.
Similarly, little benefit was seen on networking benchmarks both
localhost and between physical server/clients where other costs
dominate. It's possible that this will only be noticable on very high
speed networks.
Jesper Dangaard Brouer independently tested this with a separate
microbenchmark from
https://github.com/netoptimizer/prototype-kernel/tree/master/kernel/mm/bench
Micro-benchmarked with [1] page_bench02:
modprobe page_bench02 page_order=0 run_flags=$((2#010)) loops=$((10**8)); \
rmmod page_bench02 ; dmesg --notime | tail -n 4
Compared to baseline: 213 cycles(tsc) 53.417 ns
- against this : 184 cycles(tsc) 46.056 ns
- Saving : -29 cycles
- Very close to expected 27 cycles saving [see below [2]]
Micro benchmarking via time_bench_sample[3], we get the cost of these
operations:
time_bench: Type:for_loop Per elem: 0 cycles(tsc) 0.232 ns (step:0)
time_bench: Type:spin_lock_unlock Per elem: 33 cycles(tsc) 8.334 ns (step:0)
time_bench: Type:spin_lock_unlock_irqsave Per elem: 62 cycles(tsc) 15.607 ns (step:0)
time_bench: Type:irqsave_before_lock Per elem: 57 cycles(tsc) 14.344 ns (step:0)
time_bench: Type:spin_lock_unlock_irq Per elem: 34 cycles(tsc) 8.560 ns (step:0)
time_bench: Type:simple_irq_disable_before_lock Per elem: 37 cycles(tsc) 9.289 ns (step:0)
time_bench: Type:local_BH_disable_enable Per elem: 19 cycles(tsc) 4.920 ns (step:0)
time_bench: Type:local_IRQ_disable_enable Per elem: 7 cycles(tsc) 1.864 ns (step:0)
time_bench: Type:local_irq_save_restore Per elem: 38 cycles(tsc) 9.665 ns (step:0)
[Mel's patch removes a ^^^^^^^^^^^^^^^^] ^^^^^^^^^ expected saving - preempt cost
time_bench: Type:preempt_disable_enable Per elem: 11 cycles(tsc) 2.794 ns (step:0)
[adds a preempt ^^^^^^^^^^^^^^^^^^^^^^] ^^^^^^^^^ adds this cost
time_bench: Type:funcion_call_cost Per elem: 6 cycles(tsc) 1.689 ns (step:0)
time_bench: Type:func_ptr_call_cost Per elem: 11 cycles(tsc) 2.767 ns (step:0)
time_bench: Type:page_alloc_put Per elem: 211 cycles(tsc) 52.803 ns (step:0)
Thus, expected improvement is: 38-11 = 27 cycles.
[mgorman@techsingularity.net: s/preempt_enable_no_resched/preempt_enable/]
Link: http://lkml.kernel.org/r/20170208143128.25ahymqlyspjcixu@techsingularity.net
Link: http://lkml.kernel.org/r/20170123153906.3122-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
374ad05ab6