Originally reported here: http://lkml.iu.edu/hypermail/linux/kernel/1712.0/07072.html We're seeing unexpected CFS throttling. I wrote down detailed description with an easily reproducible case here: * https://gist.github.com/bobrik/2030ff040fad360327a5fab7a09c4ff1 TL;DR is that even though we don't burn through CFS quota during the period, throttling kicks in anyways. The larger interval between wakes up is, the higher chance of throttling we get. We're seeing this on different kernels (4.4, 4.9, 4.14) and different distros (different kernel configs). In corner cases well behaved applications that stay within quota get delayed for two periods in a row, which means 200ms latency with systemd.
This issue has been resolved by 512ac999. However 512ac999 caused another separate regression that I've submitted patches for. https://lkml.org/lkml/2019/5/17/581
My fixes have been accepted as commit de53fd7aed, and a clean-up fix of 763a9ec06c4. This should be closed as resolved.