On Thu, Mar 23, 2017 at 08:38:43AM -0700, Dave Hansen wrote:
> On 03/22/2017 01:41 AM, Aaron Lu wrote:
> > On Wed, Mar 22, 2017 at 03:33:35PM +0900, Minchan Kim wrote:
> >> On Wed, Mar 15, 2017 at 05:00:02PM +0800, Aaron Lu wrote:
> >>> Introduce a workqueue for all the free workers so that user can fine
> >>> tune how many workers can be active through sysfs interface: max_active.
> >>> More workers will normally lead to better performance, but too many can
> >>> cause severe lock contention.
> >>
> >> Let me ask a question.
> >>
> >> How well can workqueue distribute the jobs in multiple CPU?
> >
> > I would say it's good enough for my needs.
> > After all, it doesn't need many kworkers to achieve the 50% time
> > decrease: 2-4 kworkers for EP and 4-8 kworkers for EX are enough from
> > previous attched data.
>
> It's also worth noting that we'd like to *also* like to look into
> increasing how scalable freeing pages to a given zone is.
Still on EX, I restricted the allocation to be only on node 1, with
120G memory allocated there:
max_active time compared to base lock from perf
base(no parallel) 3.81s ±3.3% N/A <1%
1 3.10s ±7.7% ↓18.6% 14.76%
2 2.44s ±13.6% ↓35.9% 36.95%
4 2.07s ±13.6% ↓45.6% 59.67%
8 1.98s ±0.4% ↓48.0% 62.59%
16 2.01s ±2.4% ↓47.2% 79.62%
If we can improve the scalibility of freeing a given zone, then parallel
free will be able to achieve more.
BTW, the lock is basically pgdat->lru_lock in release_pages and
zone->lock in free_pcppages_bulk:
62.59% 62.59% [kernel.kallsyms] [k] native_queued_spin_lock_slowpath
37.17%
native_queued_spin_lock_slowpath;_raw_spin_lock_irqsave;free_pcppages_bulk;free_hot_cold_page;free_hot_cold_page_list;release_pages;free_pages_and_swap_cache;tlb_flush_mmu_free_batches;batch_free_work;process_one_work;worker_thread;kthread;ret_from_fork
25.27%
native_queued_spin_lock_slowpath;_raw_spin_lock_irqsave;release_pages;free_pages_and_swap_cache;tlb_flush_mmu_free_batches;batch_free_work;process_one_work;worker_thread;kthread;ret_from_fork
