On 03/01/2018 03:00 PM, Michal Hocko wrote:
> On Thu 01-03-18 14:28:45, Aaron Lu wrote:
>> When a page is freed back to the global pool, its buddy will be checked
>> to see if it's possible to do a merge. This requires accessing buddy's
>> page structure and that access could take a long time if it's cache cold.
>> This patch adds a prefetch to the to-be-freed page's buddy outside of
>> zone->lock in hope of accessing buddy's page structure later under
>> zone->lock will be faster. Since we *always* do buddy merging and check
>> an order-0 page's buddy to try to merge it when it goes into the main
>> allocator, the cacheline will always come in, i.e. the prefetched data
>> will never be unused.
>> In the meantime, there are two concerns:
>> 1 the prefetch could potentially evict existing cachelines, especially
>> for L1D cache since it is not huge;
>> 2 there is some additional instruction overhead, namely calculating
>> buddy pfn twice.
>> For 1, it's hard to say, this microbenchmark though shows good result but
>> the actual benefit of this patch will be workload/CPU dependant;
>> For 2, since the calculation is a XOR on two local variables, it's expected
>> in many cases that cycles spent will be offset by reduced memory latency
>> later. This is especially true for NUMA machines where multiple CPUs are
>> contending on zone->lock and the most time consuming part under zone->lock
>> is the wait of 'struct page' cacheline of the to-be-freed pages and their
>> Test with will-it-scale/page_fault1 full load:
>> kernel Broadwell(2S) Skylake(2S) Broadwell(4S) Skylake(4S)
>> v4.16-rc2+ 9034215 7971818 13667135 15677465
>> patch2/3 9536374 +5.6% 8314710 +4.3% 14070408 +3.0% 16675866 +6.4%
>> this patch 10338868 +8.4% 8544477 +2.8% 14839808 +5.5% 17155464 +2.9%
>> Note: this patch's performance improvement percent is against patch2/3.
> I am really surprised that this has such a big impact.
It's even stranger to me. Struct page is 64 bytes these days, exactly a
a cache line. Unless that changed, Intel CPUs prefetched a "buddy" cache
line (that forms an aligned 128 bytes block with the one we touch).
Which is exactly a order-0 buddy struct page! Maybe that implicit
prefetching stopped at L2 and explicit goes all the way to L1, can't
remember. Would that make such a difference? It would be nice to do some
perf tests with cache counters to see what is really going on...
> Is this a win on
> other architectures as well?
>> [changelog stole from Dave Hansen and Mel Gorman's comments]
> Please use http://lkml.kernel.org/r/<msg-id> for references because
> lkml.org is quite unstable. It would be