On Wed, Sep 26, 2018 at 12:52:22PM +0000, Daniel McGinnes wrote: > Hi, > > I plotted the charts with Percpu & SUnreclaim from /proc/meminfo. Very > interesting that it looks like Percpu increases during the test at a > similar rate MemAvailable is dropping (I plotted them on different axis so > it was easier to see). It is also noticeable that when MemAvailable > plateaus so does Percpu. > > > > So looks like sometimes Percpu increases even though nr_dying_descendant > reamins stable (I had to remove from the chart otherwise there was too > much data - but it has remained stable at 400-500 since 6 hours) - > although it does look like eventually Percpu & MemAvailable have > stabilised (although need to monitor for longer to be sure - given it > stabilised for a few hours 4 hours in, and then started increasing again) Hm, interesting. Is it current version from the Linus's tree? Anyway, let's try to figure out what's taking per-cpu memory. I've used a modified version of memleak.py from bcc (https://github.com/iovisor/bcc) to debug such things. Please, make the original memleak.py work for you, and I'll try to find a modified version for tracing the per-cpu allocator. Thanks!
