Notes from Chris Down presentation

Different types of memory - anonymous (unbacked), cache/buffers (part of unified page cache), etc.
'Reclaimability' is a factor of use.
RSS - skews towards anonymous, mapped file. A program might rely on file buffers/caches to work performantly.
Swap - increases reclaim equality and reliability of forward progress of the system, NOT emergency memory. Can reduce page cache availability.
Swap - when not present causes a fall off the cliff very quickly when memory pressure is high.
Swap - People worry that it can delay OOM. But it is a blunt tool that occurs when reclaim fails.
pte_young() gives an indication of page usage. We don't know we're OOM ahead of time.
oom killer will kill the highest RSS, not always the right choice.
Reclaim is done by kswapd in the background when resident pages goes above a threshold.
Direct reclaim is when an application has no memory available and is blocked until it is found.
Tries to reclaim the coldest pages first, but some pages might not be reclaimable - e.g. very hot cache pages, dirty pages, pages otherwise unavailable.
Hard to determine memory pressure even with scan rates - might be very efficient.
Under cgroupv2 can do:

cat /sys/fs/cgroup/system.slice/memory.pressure

Kernel measurement 'psi' - measures e.g. undesirable memory operations.
oomd - bit.ly/fboomd - Early-warning highly configurable OOM detector.
Limits don't compose very well - easy to get wrong. Preferable to use a 'protection' threshold, e.g. memory.{low,min} - won't do reclaim below min, and low only if system-wide memory contention.
Writeback threshold.
bit.ly/fbtax2 - FB tool for managing resources.
vm.overcommit_memory - Operates at VM level so not so useful.
Can get PSI from /proc/pressure/memory.
CONFIG_PSI - Get pressure measures.

Provide feedback