[lkml]   [2017]   [Apr]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [patch] mm, vmscan: avoid thrashing anon lru when free + file is low
On Tue 18-04-17 14:32:56, David Rientjes wrote:
> If the suggestion is checking
> NR_ACTIVE_ANON + NR_INACTIVE_ANON > total_high_wmark pages, it would be a
> separate heurstic to address a problem that I'm not having :) My issue is
> specifically when NR_ACTIVE_FILE + NR_INACTIVE_FILE < total_high_wmark,
> NR_ACTIVE_ANON + NR_INACTIVE_ANON is very large, but all not on this
> lruvec's evictable lrus.

Hmm, why are those pages not moved to the unevictable LRU lists?

> This is the reason why I chose lruvec_lru_size() rather than per-node
> statistics. The argument could also be made for the file lrus in the
> get_scan_count() heuristic that forces SCAN_ANON, but I have not met such
> an issue (yet). I could follow-up with that change or incorporate it into
> a v2 of this patch if you'd prefer.
> In other words, I want get_scan_count() to not force SCAN_ANON and
> fallback to SCAN_FRACT, absent other heuristics, if the amount of
> evictable anon is below a certain threshold for this lruvec. I
> arbitrarily chose SWAP_CLUSTER_MAX to be conservative, but I could easily
> compare to total_high_wmark as well, although I would consider that more
> aggressive.
> So we're in global reclaim, our file lrus are below thresholds, but we
> don't want to force SCAN_ANON for all lruvecs if there's not enough to
> reclaim from evictable anon. Do you have a suggestion for how to
> implement this logic other than this patch?

I agree that forcing SCAN_ANON without looking at the ANON lru size is
not optimal but I would rather see the same criterion for both anon and
file. get_scan_count is full of magic heuristics which tend to break for
different workloads. Let's not add another magic on top please.
Michal Hocko

 \ /
  Last update: 2017-04-19 09:05    [W:0.133 / U:4.764 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site