Re: REGRESSION: Performance regressions from switchinganon_vma->lock to mutex
From: Tim Chen
Date: Fri Jun 17 2011 - 14:22:22 EST
On Thu, 2011-06-16 at 20:58 -0700, Linus Torvalds wrote:
>
> So Tim, I'd like you to test out my first patch (that only does the
> anon_vma_clone() case) once again, but now in the cleaned-up version.
> Does this patch really make a big improvement for you? If so, this
> first step is probably worth doing regardless of the more complicated
> second step, but I'd want to really make sure it's ok, and that the
> performance improvement you saw is consistent and not a fluke.
>
> Linus
Linus,
For this patch, I've run it 10 times and got an average throughput of
104.9% compared with 2.6.39 vanilla baseline. Wide variations are seen
run to run and the difference between max and min throughput is 52% of
average value.
So to recap,
Throughput
2.6.39(vanilla) 100.0%
2.6.39+ra-patch 166.7% (+66.7%)
3.0-rc2(vanilla) 68.0% (-32%)
3.0-rc2+linus 115.7% (+15.7%)
3.0-rc2+linus+softirq 86.2% (-17.3%)
3.0-rc2+linus (v2) 104.9% (+4.9%)
The time spent in the anon_vma mutex seems to directly affect
throughput.
In one run on your patch, I got a low throughput of 90.1% vs 2.6.39
throughput. The mutex_lock occupied 15.6% of cpu.
In another run, I got a high throughput of 120.8% vs 2.6.39 throughput.
The mutex lock occupied 7.5% of cpu.
I've attached the profiles of the two runs and a 3.0-rc2 vanilla run for
your reference.
I will follow up later with numbers that has Peter's patch added.
Thanks.
Tim
----------Profiles Below-------------------------
3.0-rc2+linus(v2) run 1 (90.1% throughput vs 2.6.39)
- 15.60% exim [kernel.kallsyms] [k] __mutex_lock_common.clone.5
- __mutex_lock_common.clone.5
- 99.99% __mutex_lock_slowpath
- mutex_lock
+ 75.52% anon_vma_lock.clone.10
+ 23.88% anon_vma_clone
- 4.38% exim [kernel.kallsyms] [k] _raw_spin_lock_irqsave
- _raw_spin_lock_irqsave
+ 82.83% cpupri_set
+ 6.75% try_to_wake_up
+ 5.35% release_pages
+ 1.72% pagevec_lru_move_fn
+ 0.93% get_page_from_freelist
+ 0.51% lock_timer_base.clone.20
+ 3.22% exim [kernel.kallsyms] [k] page_fault
+ 2.62% exim [kernel.kallsyms] [k] do_raw_spin_lock
+ 2.30% exim [kernel.kallsyms] [k] mutex_unlock
+ 2.02% exim [kernel.kallsyms] [k] unmap_vmas
3.0-rc2_linus(v2) run 2 (120.8% throughput vs 2.6.39)
- 7.53% exim [kernel.kallsyms] [k] __mutex_lock_common.clone.5
- __mutex_lock_common.clone.5
- 99.99% __mutex_lock_slowpath
- mutex_lock
+ 75.99% anon_vma_lock.clone.10
+ 22.68% anon_vma_clone
+ 0.70% unlink_file_vma
- 4.15% exim [kernel.kallsyms] [k] _raw_spin_lock_irqsave
- _raw_spin_lock_irqsave
+ 83.37% cpupri_set
+ 7.06% release_pages
+ 2.74% pagevec_lru_move_fn
+ 2.18% try_to_wake_up
+ 0.99% get_page_from_freelist
+ 0.59% lock_timer_base.clone.20
+ 0.58% lock_hrtimer_base.clone.16
+ 4.06% exim [kernel.kallsyms] [k] page_fault
+ 2.33% exim [kernel.kallsyms] [k] unmap_vmas
+ 2.22% exim [kernel.kallsyms] [k] do_raw_spin_lock
+ 2.05% exim [kernel.kallsyms] [k] page_cache_get_speculative
+ 1.98% exim [kernel.kallsyms] [k] mutex_unlock
3.0-rc2 vanilla run
- 18.60% exim [kernel.kallsyms] [k] __mutex_lock_common.clone.5 â
- __mutex_lock_common.clone.5 â
- 99.99% __mutex_lock_slowpath â
- mutex_lock â
- 99.54% anon_vma_lock.clone.10 â
+ 38.99% anon_vma_clone â
+ 37.56% unlink_anon_vmas â
+ 11.92% anon_vma_fork â
+ 11.53% anon_vma_free â
- 4.03% exim [kernel.kallsyms] [k] _raw_spin_lock_irqsave â
- _raw_spin_lock_irqsave â
+ 87.25% cpupri_set â
+ 4.75% release_pages â
+ 3.68% try_to_wake_up â
+ 1.17% pagevec_lru_move_fn â
+ 0.71% get_page_from_freelist â
+ 3.00% exim [kernel.kallsyms] [k] do_raw_spin_lock â
+ 2.90% exim [kernel.kallsyms] [k] page_fault â
+ 2.25% exim [kernel.kallsyms] [k] mutex_unlock â
+ 1.82% exim [kernel.kallsyms] [k] unmap_vmas â
+ 1.62% exim [kernel.kallsyms] [k] copy_page_c â
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/