On 03/29/2018 09:42 PM, Aaron Lu wrote:
On Thu, Mar 29, 2018 at 03:19:46PM -0400, Daniel Jordan wrote:
On 03/20/2018 04:54 AM, Aaron Lu wrote:
This series is meant to improve zone->lock scalability for order 0 pages.
With will-it-scale/page_fault1 workload, on a 2 sockets Intel Skylake
server with 112 CPUs, CPU spend 80% of its time spinning on zone->lock.
Perf profile shows the most time consuming part under zone->lock is the
cache miss on "struct page", so here I'm trying to avoid those cache
misses.
I ran page_fault1 comparing 4.16-rc5 to your recent work, these four patches
plus the three others from your github branch zone_lock_rfc_v2. Out of
curiosity I also threw in another 4.16-rc5 with the pcp batch size adjusted
so high (10922 pages) that we always stay in the pcp lists and out of buddy
completely. I used your patch[*] in this last kernel.
This was on a 2-socket, 20-core broadwell server.
There were some small regressions a bit outside the noise at low process
counts (2-5) but I'm not sure they're repeatable. Anyway, it does improve
the microbenchmark across the board.
Thanks for the result.
The limited improvement is expected since lock contention only shifts,
not entirely gone. So what is interesting to see is how it performs with
v4.16-rc5 + my_zone_lock_patchset + your_lru_lock_patchset
Yep, that's 'coming soon.'
