-----Original Message----- From: Michal Hocko [mailto:[email protected]] Sent: Friday, September 15, 2017 7:50 PM To: Wang, Kemi <[email protected]> Cc: Luis R . Rodriguez <[email protected]>; Kees Cook <[email protected]>; Andrew Morton <[email protected]>; Jonathan Corbet <[email protected]>; Mel Gorman <[email protected]>; Johannes Weiner <[email protected]>; Christopher Lameter <[email protected]>; Sebastian Andrzej Siewior <[email protected]>; Vlastimil Babka <[email protected]>; Hillf Danton <[email protected]>; Dave <[email protected]>; Chen, Tim C <[email protected]>; Kleen, Andi <[email protected]>; Jesper Dangaard Brouer <[email protected]>; Huang, Ying <[email protected]>; Lu, Aaron <[email protected]>; Proc sysctl <[email protected]>; Linux MM <[email protected]>; Linux Kernel <[email protected]> Subject: Re: [PATCH 1/3] mm, sysctl: make VM stats configurable
On Fri 15-09-17 17:23:24, Kemi Wang wrote: > This patch adds a tunable interface that allows VM stats configurable, as > suggested by Dave Hansen and Ying Huang. > > When performance becomes a bottleneck and you can tolerate some possible > tool breakage and some decreased counter precision (e.g. numa counter), you > can do: > echo [C|c]oarse > /proc/sys/vm/vmstat_mode > > When performance is not a bottleneck and you want all tooling to work, you > can do: > echo [S|s]trict > /proc/sys/vm/vmstat_mode > > We recommend automatic detection of virtual memory statistics by system, > this is also system default configuration, you can do: > echo [A|a]uto > /proc/sys/vm/vmstat_mode > > The next patch handles numa statistics distinctively based-on different VM > stats mode. I would just merge this with the second patch so that it is clear how those modes are implemented. I am also wondering why cannot we have a much simpler interface and implementation to enable/disable numa stats (btw. sysctl_vm_numa_stats would be more descriptive IMHO). The motivation is that we propose a general tunable interface for VM stats. This would be more scalable, since we don't have to add an individual Interface for each type of counter that can be configurable. In the second patch, NUMA stats, as an example, can benefit for that. If you still hold your idea, I don't mind to merge them together. -- Michal Hocko SUSE Labs
