On Thu, Oct 13, 2016 at 12:59:14PM +0200, Vlastimil Babka wrote:
> On 10/13/2016 10:08 AM, js1...@gmail.com wrote:
> >From: Joonsoo Kim
> >
> >Allocation/free pattern is usually sequantial. If they are freed to
> >the buddy list, they can be coalesced. However, we first keep these freed
> >pages at the pcp list and try to reuse them until threshold is reached
> >so we don't have enough chance to get a high order freepage. This reusing
> >would provide us some performance advantages since we don't need to
> >get the zone lock and we don't pay the cost to check buddy merging.
> >But, less fragmentation and more high order freepage would compensate
> >this overhead in other ways. First, we would trigger less direct
> >compaction which has high overhead. And, there are usecases that uses
> >high order page to boost their performance.
> >
> >Instantly resuing freed page seems to provide us computational benefit
> >but the other affects more precious things like as I/O performance and
> >memory consumption so I think that it's a good idea to weight
> >later advantage more.
>
> Again, there's also cache hotness to consider. And whether the
> sequential pattern is still real on a system with higher uptime.
> Should be possible to evaluate with tracepoints?
I answered this in previous e-mail. Anyway, we should evaluate
cache-effect. tracepoint or perf's cache event would show some
evidence. I will do it soon and report again.
Thanks.
