Re: [RFC PATCH v1 13/13] mm: splice local lists onto the front of the LRU

[Aaron Lu]

On Wed, Jan 31, 2018 at 06:04:13PM -0500, daniel.m.jordan@xxxxxxxxxx wrote:
> Now that release_pages is scaling better with concurrent removals from
> the LRU, the performance results (included below) showed increased
> contention on lru_lock in the add-to-LRU path.
> 
> To alleviate some of this contention, do more work outside the LRU lock.
> Prepare a local list of pages to be spliced onto the front of the LRU,
> including setting PageLRU in each page, before taking lru_lock.  Since
> other threads use this page flag in certain checks outside lru_lock,
> ensure each page's LRU links have been properly initialized before
> setting the flag, and use memory barriers accordingly.
> 
> Performance Results
> 
> This is a will-it-scale run of page_fault1 using 4 different kernels.
> 
>             kernel     kern #
> 
>           4.15-rc2          1
>   large-zone-batch          2
>      lru-lock-base          3
>    lru-lock-splice          4
> 
> Each kernel builds on the last.  The first is a baseline, the second
> makes zone->lock more scalable by increasing an order-0 per-cpu
> pagelist's 'batch' and 'high' values to 310 and 1860 respectively

Since the purpose of the patchset is to optimize lru_lock, you may
consider adjusting pcp->high to be >= 32768(page_fault1's test size is
128M = 32768 pages). That should eliminate zone->lock contention
entirely.

> (courtesy of Aaron Lu's patch), the third scales lru_lock without
> splicing pages (the previous patch in this series), and the fourth adds
> page splicing (this patch).
> 
> N tasks mmap, fault, and munmap anonymous pages in a loop until the test
> time has elapsed.