Re: [PATCH] powerpc/mm: Fix RECLAIM_DISTANCE

[Balbir Singh <bsingharora@gmail.com>]

On Wed, Jan 25, 2017 at 03:58:22PM +1100, Gavin Shan wrote:
> On Wed, Jan 25, 2017 at 09:27:44AM +0530, Balbir Singh wrote:
> >On Tue, Jan 24, 2017 at 10:32:28AM +1100, Gavin Shan wrote:
> >> When @node_reclaim_mode ("/proc/sys/vm/zone_reclaim_mode") is enabled,
> >> the nodes in the specified distance (< RECLAIM_DISTANCE) to the preferred
> >> one will be checked for page direct reclaim in the fast path, as below
> >> function call chain indicates. Currently, RECLAIM_DISTANCE is set to 10,
> >> equal to LOCAL_DISTANCE. It means no nodes, including the preferred one,
> >> don't match the conditions. So no nodes are checked for direct reclaim
> >> in the fast path.
> >> 
> >>    __alloc_pages_nodemask
> >>    get_page_from_freelist
> >>    zone_allows_reclaim
> >> 
> >> This fixes it by setting RECLAIM_DISTANCE to 30. With it, the preferred
> >> node and its directly adjacent nodes will be checked for page direct
> >> reclaim. The comments explaining RECLAIM_DISTANCE is out of date. This
> >> updates and makes it correct.
> >> 
> >> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
> >> ---
> >
> >I spoke about this at length with Anton and the larger kernel team.
> >We need performance data before we can commit to the change. Do we
> >benchmarks to show that the change does not introduce regression
> >w.r.t runtime cost?
> >
> 
> Thanks for review. I just found the problem when studying the code
> last year. It sounds reasonable not to rely on the slow path for page
> reclaim if the fast path can reclaim enough pages. From this point,
> I believe the performance should be improved. In the meanwhile, the
> page cache/buffer could be released, as part of the output of page
> reclaim. It's going to affect fs's performance for sure. So do you
> have recommended test examples to measure the improved performance
> because of this?
>

Anton suggested that NUMA distances in powerpc mattered and hurted
performance without this setting. We need to validate to see if this
is still true. A simple way to start would be benchmarking

Balbir Singh.