Part 2 · 1 chapters · ~8 min

Scheduling: CFS and EEVDF

Tasks and the run queue, scheduling classes (stop, deadline, real-time, fair, idle), virtual runtime and nice weights, CFS and its red-black tree, EEVDF since Linux 6.6, preemption and context switches, CPU affinity and load balancing, cgroup CPU controls and throttling, and sched_ext.

Who runs next

class	policies	use
deadline	SCHED_DEADLINE	tasks with runtime, period and deadline guarantees
real-time	SCHED_FIFO, SCHED_RR	audio, control loops; can starve everything below
fair	SCHED_NORMAL, SCHED_BATCH	almost every process (CFS, now EEVDF)
idle	SCHED_IDLE	runs only when nothing else wants the CPU

code

chrt -p $(pgrep -o postgres)                 # current policy and priority
taskset -cp 0-3 $(pgrep -o node)             # pin to CPUs 0-3
cat /proc/$(pgrep -o node)/sched | head      # se.vruntime, nr_switches, voluntary vs involuntary
cat /sys/fs/cgroup/<group>/cpu.stat          # nr_throttled, throttled_usec: CPU limits biting

CPU limits and latency: a container with a CPU limit uses the CFS bandwidth controller (cpu.max: quota per period). A multi-threaded runtime can spend its quota early in a 100 ms period and be throttled for the rest, adding tail latency even at low average CPU. Watch nr_throttled. sched_ext (merged in 6.12) lets schedulers be written as eBPF programs.

SCHEDULING: FROM CFS TO EEVDF

fair shares of CPU, and who runs next

swipe the figure sideways, or tap expand for full screen

1/4

per-CPU queues

Each CPU has its own run queue, so scheduling decisions are local and cheap; load balancing moves tasks between CPUs periodically.

a queue per CPUbalanced periodically