On Sat, Jan 23, 2021 at 4:29 AM Zbigniew Jędrzejewski-Szmek
<zbyszek(a)in.waw.pl> wrote:
(One possible direction: one thing I want to explore next is using zram
or zwap based on whether the machine has a physical swap device. Maybe
such a language would be useful then — with additional variables
specifying e.g. the physical swap size…)
What about setting vm.swappiness = 120?
When set to 100, the bias for reclaiming anonymous pages and file
pages is about equal. Setting it lower is predicated on (a) older
kernels and (b) spinning drives where the cost for page out and page
in is higher than dropping a file page and only reading it back in.
With zram based swap, eviction and reclaim of anon pages is
unquestionably a lot cheaper now, and even cheaper than the cost of
reading in a file page. I'm even thinking it could be pushed higher
than 120. I don't think there's a way to make this smart enough to
scale this to the swap backing storage performance, which is what we
really want. Hence 120 is a compromise in case there's also disk based
swap.
A down the road enhancement might be, if no disk based swap is
detected, push this to 190. This would also allow some time to get
some feedback with it set to 120 before pushing harder.
--
Chris Murphy