On Mon, Dec 07, 2020 at 11:46:22AM -0800, Roman Gushchin wrote: > On Sun, Dec 06, 2020 at 06:14:48PM +0800, Muchun Song wrote: > > the global and per-node counters are stored in pages, however memcg > > and lruvec counters are stored in bytes. This scheme looks weird. > > So convert all vmstat slab counters to bytes. > > There is a reason for this weird scheme: > percpu caches (see struct per_cpu_nodestat) are s8, so counting in bytes > will lead to overfills. Switching to s32 can lead to an increase in > the cache thrashing, especially on small machines. JFYI: I've tried to convert all slab counters to bytes and change those s8 percpu batches to s32 about a year ago. Here is a link to that thread: https://patchwork.kernel.org/project/linux-mm/patch/20191018002820.307763-3-guro@xxxxxx/ Thanks!
