Commit 1a61ab8038e72 ("mm: memcontrol: replace zone summing with lruvec_page_state()") has made lruvec_page_state to use per-cpu counters instead of calculating it directly from lru_zone_size with an idea that this would be more effective.
Tim has reported that this is not really the case for their database benchmark which is showing an opposite results where lruvec_page_state is taking up a huge chunk of CPU cycles (about 25% of the system time which is roughly 7% of total cpu cycles) on 5.3 kernels. The workload is running on a larger machine (96cpus), it has many cgroups (500) and it is heavily direct reclaim bound.
Tim Chen said:
: The problem can also be reproduced by running simple multi-threaded : pmbench benchmark with a fast Optane SSD swap (see profile below). : : : 6.15% 3.08% pmbench [kernel.vmlinux] [k] lruvec_lru_size : | : |--3.07%--lruvec_lru_size : | | : | |--2.11%--cpumask_next : | | | : | | --1.66%--find_next_bit : | | : | --0.57%--call_function_interrupt : | | : | --0.55%--smp_call_function_interrupt : | : |--1.59%--0x441f0fc3d009 : | _ops_rdtsc_init_base_freq : | access_histogram : | page_fault : | __do_page_fault : | handle_mm_fault : | __handle_mm_fault : | | : | --1.54%--do_swap_page : | swapin_readahead : | swap_cluster_readahead : | | : | --1.53%--read_swap_cache_async : | __read_swap_cache_async : | alloc_pages_vma : | __alloc_pages_nodemask : | __alloc_pages_slowpath : | try_to_free_pages : | do_try_to_free_pages : | shrink_node : | shrink_node_memcg : | | : | |--0.77%--lruvec_lru_size : | | : | --0.76%--inactive_list_is_low : | | : | --0.76%--lruvec_lru_size : | : --1.50%--measure_read : page_fault : __do_page_fault : handle_mm_fault : __handle_mm_fault : do_swap_page : swapin_readahead : swap_cluster_readahead : | : --1.48%--read_swap_cache_async : __read_swap_cache_async : alloc_pages_vma : __alloc_pages_nodemask : __alloc_pages_slowpath : try_to_free_pages : do_try_to_free_pages : shrink_node : shrink_node_memcg : | : |--0.75%--inactive_list_is_low : | | : | --0.75%--lruvec_lru_size : | : --0.73%--lruvec_lru_size
The likely culprit is the cache traffic the lruvec_page_state_local generates. Dave Hansen says:
: I was thinking purely of the cache footprint. If it's reading : pn->lruvec_stat_local->count[idx] is three separate cachelines, so 192 : bytes of cache *96 CPUs = 18k of data, mostly read-only. 1 cgroup would : be 18k of data for the whole system and the caching would be pretty : efficient and all 18k would probably survive a tight page fault loop in : the L1. 500 cgroups would be ~90k of data per CPU thread which doesn't : fit in the L1 and probably wouldn't survive a tight page fault loop if : both logical threads were banging on different cgroups. : : It's just a theory, but it's why I noted the number of cgroups when I : initially saw this show up in profiles
Fix the regression by partially reverting the said commit and calculate the lru size explicitly.
Link: http://email@example.com Fixes: 1a61ab8038e72 ("mm: memcontrol: replace zone summing with lruvec_page_state()")
b11edebbc967 mm: memcg: get number of pages on the LRU list in memcgroup base on lru_zone_size
mm/vmscan.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)