On Tue, Aug 6, 2019 at 3:35 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Tue 06-08-19 03:19:00, Yafang Shao wrote: > > In the node reclaim, may_shrinkslab is 0 by default, > > hence shrink_slab will never be performed in it. > > While shrik_slab should be performed if the relcaimable slab is over > > min slab limit. > > > > Add scan_control::no_pagecache so shrink_node can decide to reclaim page > > cache, slab, or both as dictated by min_unmapped_pages and min_slab_pages. > > shrink_node will do at least one of the two because otherwise node_reclaim > > returns early. > > > > __node_reclaim can detect when enough slab has been reclaimed because > > sc.reclaim_state.reclaimed_slab will tell us how many pages are > > reclaimed in shrink slab. > > > > This issue is very easy to produce, first you continuously cat a random > > non-exist file to produce more and more dentry, then you read big file > > to produce page cache. And finally you will find that the denty will > > never be shrunk in node reclaim (they can only be shrunk in kswapd until > > the watermark is reached). > > > > Regarding vm.zone_reclaim_mode, we always set it to zero to disable node > > reclaim. Someone may prefer to enable it if their different workloads work > > on different nodes. > > Considering that this is a long term behavior of a rarely used node > reclaim I would rather not touch it unless some _real_ workload suffers > from this behavior. Or is there any reason to fix this even though there > is no evidence of real workloads suffering from the current behavior? > -- When we do performance tuning on some workloads(especially if this workload is NUMA sensitive), sometimes we may enable it on our test environment and then do some benchmark to dicide whether or not applying it on the production envrioment. Although the result is not good enough as expected, it is really a performance tuning knob. Thanks Yafang
