Hi all,
Some time ago KOSAKI Motohiro noticed[1] that vmevent might be very inaccurate (up to 2GB inaccuracy on a very large machines) since per CPU stats synchronization happens either on time basis or when we hit stat thresholds.
KOSAKI also told that perf API might be a good inspirations for further improvements, but I must admit I didn't fully get the idea, although I'm open to investigate this route too, but I guess it needs a bit more explanations.
Also note that this is just an RFC, I just show some ideas and wonder how you feel about it. Since we now use memory pressure factor bolted into the reclaimer code path, we don't desperately need the accurate stats, but it's still nice thing to have/fix.
Anyway, here we take two approaches:
- Asynchronously sum vm_stat diffs and global stats. This is very similar to what we already have for per-zone stats, implemented in zone_page_state_snapshot(). The values still could be inaccurate, but overall this makes things better;
- Implement configurable per CPU vmstat thresholds. This is much more powerful tool to get accurate statistics, but it comes with a price: it might cause some performance penalty as we'd update global stats more frequently (in a fast path), so users have to be careful.
The two items are independent, so we might implement one or another, or both, or none, if desired. ;-)
Thanks, Anton.
p.s. Note that the patches are against my vmevent tree, i.e.:
git://git.infradead.org/users/cbou/linux-vmevent.git
[1] http://lkml.indiana.edu/hypermail/linux/kernel/1205.1/00062.html