Hard
CFS bandwidth throttling with low average CPU usage
A containerized service has a cgroup CPU quota of 2 cores (cpu.cfs_quota_us=200000, cpu.cfs_period_us=100000, i.e. 200ms of CPU time allowed per 100ms period). Dashboards show its average CPU usage over any 1-minute window is only ~40% of that quota, yet cgroup metrics (nr_throttled / throttled_time) show it being throttled dozens of times per minute, and request p99 latency has periodic spikes. How can a workload with low average usage still get throttled this often?
33 of 53 in Operating Systems