kswapd pegging cpu

[kevin gunn]

while debugging bug 1491566
On arale, rc-proposed circa sept 14
I had launched phone, messaging, browser (on cnn.com), camera (taken pics), gallery (edit mode of pic), clock, ubuntu-system-settings, cut-the-rope, dash (of course) - toggling between them in the spread, just selecting them, mostly clock, browser, gallery
and at the same time running the phone/text script as found in bug 1491566 (modified to text, call, then text 5 times)

I just happened to be running top and then noticed process
kswapd
would run & peg cpu intermittently, sometimes it would peg the cpu for just 1 second or so, but then sometimes it would appear to run at 95%+ for like 10 seconds or more.
It didn't seem to follow a particular pattern, but i could definitely tell it was effecting the ui (stalling)

ogra indicated our swappiness should be very low.
so i did sudo sysctl vm.swappiness=10 as instructed, however as i continued testing, i continued to see kswapd continue to run and peg the cpu at about the same frequency.

[kevin gunn (kgunn72)]

just adding, i'm not an anomaly jibel indicates he sees this quite regularly in testing

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Jean-Baptiste Lallement (jibel)]

I think kswapd is a necessary evil and the system can hardly be prevented of swapping pages in and out. kswapd activity is generally a symptom of another problem.

I noticed that when kswapd runs for several seconds, it generates lot of IO and slows down the system drastically. It can be reproduced with dd and creating large files. The system slow down due IO load is also observed when apport writes a crash report. Is there a way to tune the filesystem and improve disk access?

Secondly, when the system swaps heavily it means that some process is using lot of memory. I looked at the memory consumption and reported bug 1497924 for the memory usage of oxide and apps depending on it (webbrowser and webapps) Fixing it might improve the situation of memory consumption and reduce the activity of kswapd.

Finally, setting swappiness to 1 *seems* to improve the situation a little bit but it's just an impression and I don't have numbers to prove if it's better one way or the other. Certainly the default of 100 has a negative impact on performance.

[Pat McGowan (pat-mcgowan)]

Not sure what to do with this

[Łukasz Zemczak (sil2100)]

I suppose we might consider discussing changing the default vm.swappiness value to something smaller, but I guess someone would have to do some testing and comparison so we have actual numbers on how it affects system performance. Not sure if I would consider it a hotfix for OTA-8.5 though, not enough research done on this.

[Launchpad Janitor (janitor)]

[Expired for lxc-android-config (Ubuntu) because there has been no activity for 60 days.]