It appears that kernels newer than 4.1 have kswapd-related bug resulting in high CPU usage. CPU 100% usage could last for several minutes or several days, with CPU being busy entirely with serving kswapd. It happens usually after server being mostly idle, sometimes after days, sometimes after weeks of uptime. But the issue appears much sooner if the machine is loaded with something like building a kernel.
Here are the graphs of CPU load: first, second. Perf top output is here as well.
To find the cause of this problem I've started with the fact that the issue appeared after 4.1 kernel update. Then I performed longterm test of 3.18, and discovered that 3.18 is unaffected by this bug. Then I did some tests of 4.0 to confirm that this version behaves well too.
Then I performed git bisect from tag v4.0 to v4.1-rc1 and found exact commits that seem to be reason of high CPU usage.
The first really "bad" commit is 79553da293d38d63097278de13e28a3b371f43c1. 2 previous commits cause weird behavior as well resulting in kswapd consuming more CPU than unaffected kernels, but not that much as the commit pointed above. I believe those commits are related to the same mm tree merge.
I tried to add transparent_hugepage=never to kernel boot parameters, but it did not change anything. Changing allocator to SLAB from SLUB alters behavior and makes CPU load lower, but don't solve a problem at all.
Here is kernel bugzilla bugreport as well.
Ideas?