> -----邮件原件----- > 发件人: Michal Hocko [mailto:mhocko@xxxxxxxxxx] > 发送时间: 2018年3月23日 18:09 > 收件人: Li,Rongqing <lirongqing@xxxxxxxxx> > 抄送: linux-kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; > cgroups@xxxxxxxxxxxxxxx; hannes@xxxxxxxxxxx; Andrey Ryabinin > <aryabinin@xxxxxxxxxxxxx> > 主题: Re: 答复: 答复: [PATCH] mm/memcontrol.c: speed up to force empty > a memory cgroup > > On Fri 23-03-18 02:58:36, Li,Rongqing wrote: > > > > > > > -----邮件原件----- > > > 发件人: linux-kernel-owner@xxxxxxxxxxxxxxx > > > [mailto:linux-kernel-owner@xxxxxxxxxxxxxxx] 代表 Li,Rongqing > > > 发送时间: 2018年3月19日 18:52 > > > 收件人: Michal Hocko <mhocko@xxxxxxxxxx> > > > 抄送: linux-kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; > > > cgroups@xxxxxxxxxxxxxxx; hannes@xxxxxxxxxxx; Andrey Ryabinin > > > <aryabinin@xxxxxxxxxxxxx> > > > 主题: 答复: 答复: [PATCH] mm/memcontrol.c: speed up to force > empty a > > > memory cgroup > > > > > > > > > > > > > -----邮件原件----- > > > > 发件人: Michal Hocko [mailto:mhocko@xxxxxxxxxx] > > > > 发送时间: 2018年3月19日 18:38 > > > > 收件人: Li,Rongqing <lirongqing@xxxxxxxxx> > > > > 抄送: linux-kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; > > > > cgroups@xxxxxxxxxxxxxxx; hannes@xxxxxxxxxxx; Andrey Ryabinin > > > > <aryabinin@xxxxxxxxxxxxx> > > > > 主题: Re: 答复: [PATCH] mm/memcontrol.c: speed up to force empty > a > > > memory > > > > cgroup > > > > > > > > On Mon 19-03-18 10:00:41, Li,Rongqing wrote: > > > > > > > > > > > > > > > > -----邮件原件----- > > > > > > 发件人: Michal Hocko [mailto:mhocko@xxxxxxxxxx] > > > > > > 发送时间: 2018年3月19日 16:54 > > > > > > 收件人: Li,Rongqing <lirongqing@xxxxxxxxx> > > > > > > 抄送: linux-kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; > > > > > > cgroups@xxxxxxxxxxxxxxx; hannes@xxxxxxxxxxx; Andrey Ryabinin > > > > > > <aryabinin@xxxxxxxxxxxxx> > > > > > > 主题: Re: [PATCH] mm/memcontrol.c: speed up to force empty a > > > > memory > > > > > > cgroup > > > > > > > > > > > > On Mon 19-03-18 16:29:30, Li RongQing wrote: > > > > > > > mem_cgroup_force_empty() tries to free only 32 > > > > (SWAP_CLUSTER_MAX) > > > > > > > pages on each iteration, if a memory cgroup has lots of page > > > > > > > cache, it will take many iterations to empty all page cache, > > > > > > > so increase the reclaimed number per iteration to speed it > > > > > > > up. same as in > > > > > > > mem_cgroup_resize_limit() > > > > > > > > > > > > > > a simple test show: > > > > > > > > > > > > > > $dd if=aaa of=bbb bs=1k count=3886080 > > > > > > > $rm -f bbb > > > > > > > $time echo > > > > 100000000 >/cgroup/memory/test/memory.limit_in_bytes > > > > > > > > > > > > > > Before: 0m0.252s ===> after: 0m0.178s > > > > > > > > > > > > Andrey was proposing something similar [1]. My main objection > > > > > > was that his approach might lead to over-reclaim. Your > > > > > > approach is more conservative because it just increases the > > > > > > batch size. The size is still rather arbitrary. Same as > > > > > > SWAP_CLUSTER_MAX but that one is a commonly used unit of > reclaim in the MM code. > > > > > > > > > > > > I would be really curious about more detailed explanation why > > > > > > having a larger batch yields to a better performance because > > > > > > we are doingg SWAP_CLUSTER_MAX batches at the lower reclaim > > > > > > level > > > anyway. > > > > > > > > > > > > > > > > Although SWAP_CLUSTER_MAX is used at the lower level, but the > > > > > call stack of try_to_free_mem_cgroup_pages is too long, increase > > > > > the nr_to_reclaim can reduce times of calling > > > > > function[do_try_to_free_pages, shrink_zones, hrink_node ] > > > > > > > > > > mem_cgroup_resize_limit > > > > > --->try_to_free_mem_cgroup_pages: .nr_to_reclaim = max(1024, > > > > > --->SWAP_CLUSTER_MAX), > > > > > ---> do_try_to_free_pages > > > > > ---> shrink_zones > > > > > --->shrink_node > > > > > ---> shrink_node_memcg > > > > > ---> shrink_list <-------loop will happen in this > place > > > > [times=1024/32] > > > > > ---> shrink_page_list > > > > > > > > Can you actually measure this to be the culprit. Because we should > > > > rethink our call path if it is too complicated/deep to perform well. > > > > Adding arbitrary batch sizes doesn't sound like a good way to go to me. > > > > > > Ok, I will try > > > > > http://pasted.co/4edbcfff > > > > This is result from ftrace graph, it maybe prove that the deep call > > path leads to low performance. > > Does it? Let's have a look at the condensed output: > 6) | try_to_free_mem_cgroup_pages() { > 6) | mem_cgroup_select_victim_node() { > 6) 0.320 us | mem_cgroup_node_nr_lru_pages(); > 6) 0.151 us | mem_cgroup_node_nr_lru_pages(); > 6) 2.190 us | } > 6) | do_try_to_free_pages() { > 6) | shrink_node() { > 6) | shrink_node_memcg() { > 6) | shrink_inactive_list() { > 6) + 23.131 us | shrink_page_list(); > 6) + 33.960 us | } > 6) + 39.203 us | } > 6) | shrink_slab() { > 6) + 72.955 us | } > 6) ! 116.529 us | } > 6) | shrink_node() { > 6) 0.050 us | mem_cgroup_iter(); > 6) 0.035 us | mem_cgroup_low(); > 6) | shrink_node_memcg() { > 6) 3.955 us | } > 6) | shrink_slab() { > 6) + 54.296 us | } > 6) + 61.502 us | } > 6) ! 185.020 us | } > 6) ! 188.165 us | } > > try_to_free_mem_cgroup_pages is the full memcg reclaim path taking > 188,165 us. The pure reclaim path is shrink_node and that took 116+61 = > 177 us. > So we have 11us spent on the way. Is this really making such a difference? > How does the profile look when we do larger batches? > > > And when increase reclaiming page in try_to_free_mem_cgroup_pages, it > > can reduce calling of shrink_slab, which save times, in my cases, page > > caches occupy most memory, slab is little, but shrink_slab will be > > called everytime > > OK, that makes more sense! shrink_slab is clearly visible here. It is more > expensive than the page reclaim. This is something to look into. > shrink_slab() { 6) 0.175 us | down_read_trylock(); 6) | super_cache_count() { 6) 0.642 us | list_lru_count_one(); 6) 0.587 us | list_lru_count_one(); 6) 3.740 us | } 6) | super_cache_count() { 6) 0.625 us | list_lru_count_one(); 6) 0.485 us | list_lru_count_one(); 6) 2.145 us | } 6) | super_cache_count() { 6) 0.333 us | list_lru_count_one(); 6) 0.334 us | list_lru_count_one(); 6) 2.109 us | } 6) | super_cache_count() { 6) 0.062 us | list_lru_count_one(); 6) 0.188 us | list_lru_count_one(); 6) 1.216 us | } 6) | super_cache_count() { 6) 0.217 us | list_lru_count_one(); 6) 0.056 us | list_lru_count_one(); 6) 1.282 us | } 6) | super_cache_count() { 6) 0.204 us | list_lru_count_one(); 6) 0.205 us | list_lru_count_one(); 6) 1.237 us | } 6) | super_cache_count() { 6) 0.596 us | list_lru_count_one(); 6) 0.493 us | list_lru_count_one(); 6) 2.140 us | } 6) | super_cache_count() { 6) 0.130 us | list_lru_count_one(); 6) 0.056 us | list_lru_count_one(); 6) 1.260 us | } 6) | super_cache_count() { 6) 0.385 us | list_lru_count_one(); 6) 0.054 us | list_lru_count_one(); 6) 1.186 us | } 6) | super_cache_count() { 6) 0.304 us | list_lru_count_one(); 6) 0.286 us | list_lru_count_one(); 6) 1.550 us | } 6) | super_cache_count() { 6) 0.230 us | list_lru_count_one(); 6) 0.128 us | list_lru_count_one(); 6) 1.408 us | } 6) | super_cache_count() { 6) 0.392 us | list_lru_count_one(); 6) 0.132 us | list_lru_count_one(); 6) 1.694 us | } 6) | super_cache_count() { 6) 0.257 us | list_lru_count_one(); 6) 0.258 us | list_lru_count_one(); 6) 1.510 us | } 6) | super_cache_count() { 6) 0.132 us | list_lru_count_one(); 6) 0.132 us | list_lru_count_one(); 6) 1.361 us | } 6) | super_cache_count() { 6) 0.130 us | list_lru_count_one(); 6) 0.130 us | list_lru_count_one(); 6) 1.246 us | } 6) | count_shadow_nodes() { 6) 0.203 us | list_lru_count_one(); 6) 0.042 us | mem_cgroup_node_nr_lru_pages(); 6) 1.131 us | } 6) | super_cache_count() { 6) 0.202 us | list_lru_count_one(); 6) 0.056 us | list_lru_count_one(); 6) 1.115 us | } 6) | super_cache_count() { 6) 0.055 us | list_lru_count_one(); 6) 0.107 us | list_lru_count_one(); 6) 0.958 us | } 6) | super_cache_count() { 6) 0.147 us | list_lru_count_one(); 6) 0.150 us | list_lru_count_one(); 6) 1.474 us | } 6) | super_cache_count() { 6) 0.491 us | list_lru_count_one(); 6) 0.485 us | list_lru_count_one(); 6) 2.569 us | } 6) | super_cache_count() { 6) 0.605 us | list_lru_count_one(); 6) 0.590 us | list_lru_count_one(); 6) 2.136 us | } 6) | super_cache_count() { 6) 0.572 us | list_lru_count_one(); 6) 0.418 us | list_lru_count_one(); 6) 1.914 us | } 6) | super_cache_count() { 6) 0.428 us | list_lru_count_one(); 6) 0.358 us | list_lru_count_one(); 6) 2.073 us | } /* super_cache_count */ 6) | super_cache_count() { 6) 0.422 us | list_lru_count_one(); 6) 0.433 us | list_lru_count_one(); 6) 1.604 us | } 6) | super_cache_count() { 6) 0.532 us | list_lru_count_one(); 6) 0.280 us | list_lru_count_one(); 6) 1.523 us | } 6) | super_cache_count() { 6) 0.422 us | list_lru_count_one(); 6) 0.574 us | list_lru_count_one(); 6) 1.554 us | } 6) | super_cache_count() { 6) 0.565 us | list_lru_count_one(); 6) 0.587 us | list_lru_count_one(); 6) 1.878 us | } 6) | super_cache_count() { 6) 0.563 us | list_lru_count_one(); 6) 0.558 us | list_lru_count_one(); 6) 1.949 us | } 6) | super_cache_count() { 6) 0.468 us | list_lru_count_one(); 6) 0.476 us | list_lru_count_one(); 6) 2.149 us | } 6) | super_cache_count() { 6) 0.443 us | list_lru_count_one(); 6) 0.483 us | list_lru_count_one(); 6) 2.283 us | } 6) | super_cache_count() { 6) 0.332 us | list_lru_count_one(); 6) 0.228 us | list_lru_count_one(); 6) 1.307 us | } 6) | super_cache_count() { 6) 0.532 us | list_lru_count_one(); 6) 0.367 us | list_lru_count_one(); 6) 1.956 us | } 6) 0.036 us | up_read(); 6) 0.038 us | _cond_resched(); 6) + 72.955 us | } shrink_slab does not reclaim any memory, but take lots of time to count lru maybe we can use the returning of shrink_slab to control if next shrink_slab should be called? Or define a slight list_lru_empty to check if sb->s_dentry_lru is empty before calling list_lru_shrink_count, like below diff --git a/fs/super.c b/fs/super.c index 672538ca9831..954c22338833 100644 --- a/fs/super.c +++ b/fs/super.c @@ -130,8 +130,10 @@ static unsigned long super_cache_count(struct shrinker *shrink, if (sb->s_op && sb->s_op->nr_cached_objects) total_objects = sb->s_op->nr_cached_objects(sb, sc); - total_objects += list_lru_shrink_count(&sb->s_dentry_lru, sc); - total_objects += list_lru_shrink_count(&sb->s_inode_lru, sc); + if (!list_lru_empty(sb->s_dentry_lru)) + total_objects += list_lru_shrink_count(&sb->s_dentry_lru, sc); + if (!list_lru_empty(sb->s_inode_lru)) + total_objects += list_lru_shrink_count(&sb->s_inode_lru, sc); total_objects = vfs_pressure_ratio(total_objects); return total_objects; -RongQing > Thanks! > -- > Michal Hocko > SUSE Labs
