Linux Process Scheduling & CPU Affinity: Strategies, Commands, cgroup Guide

Scheduling Policies

Linux can run multiple processes on a single CPU by interleaving their execution. The kernel scheduler decides which process runs at any moment, balancing fast decision making, fair CPU time distribution (with higher priority processes receiving more time and preempting lower ones), responsive interactive applications, and predictable scalability under varied workloads.

Process Priorities

Scheduling decisions are based on the policy and priority assigned to each thread or process. Policies are divided into real‑time and non‑real‑time categories.

Available Scheduling Policies (RHEL)

Real‑time

SCHED_FIFO: First‑in‑first‑out without time slicing; a process runs until it blocks on I/O or is preempted by a higher‑priority process.

SCHED_RR: Round‑robin with time slices; processes of equal priority share CPU time in a cyclic order.

Non‑real‑time

SCHED_NORMAL (OTHER): Default policy for most Linux processes.

SCHED_BATCH: Suitable for batch processing workloads.

SCHED_IDLE: Allows low‑priority applications to run when the CPU is idle.

CFS Scheduler

Since kernel 2.6.23, the Completely Fair Scheduler (CFS) is the default. It uses a red‑black tree and virtual time to select the process that has waited the longest, granting it CPU time while its virtual time decreases.

Deadline Scheduler

RHEL 8 introduces SCHED_DEADLINE for real‑time systems, guaranteeing execution based on three parameters: period, deadline, and runtime (worst‑case execution time).

Period: The interval at which a task repeats (e.g., 16 ms for 60 fps video processing).

Deadline: The maximum time by which the task must finish.

Runtime: The maximum CPU time the task may consume.

All time units are in nanoseconds; for example, a task with a 16.6 ms period must receive 5 ms of CPU time, which must be completed within a 10 ms deadline.

Changing Scheduling Options via Command Line

Administrators can use chrt to view or set a process's policy and priority. When launching a new program with chrt, the default policy is SCHED_RR if none is specified.

Policy options:

-b  SCHED_BATCH
-f  SCHED_FIFO
-i  SCHED_IDLE
-o  SCHED_NORMAL (OTHER)
-r  SCHED_RR
-d  SCHED_DEADLINE

Example: assigning a policy and priority to a new process.

Changing Scheduling Options via systemd

When defining a service, the [Service] section can specify:

CPUSchedulingPolicy: Sets the service's CPU scheduling policy (other, batch, idle, fifo, rr). Deadline is not currently supported.

CPUSchedulingPriority: Sets the priority (1‑99 for real‑time policies).

After editing, reload systemd with systemctl daemon-reload and restart the service.

CPU Affinity

Pinning processes to specific CPUs can improve cache locality and overall performance. Systemd services can set CPUAffinity in the [Service] section, providing a space‑separated list of CPU indices (e.g., "0 1").

Managing CPU Affinity with cgroup

On NUMA systems, binding a process to CPUs within the same node reduces memory access latency. The cpuset cgroup controller allows specifying which CPUs and memory nodes a group of tasks may use.

Key files in /sys/fs/cgroup/cpuset:

cpuset.cpus – List of CPUs the cgroup can access (e.g., "0-3").

cpuset.mems – List of NUMA memory nodes the cgroup can access.

Use the tuna tool to view CPU bindings:

yum install tuna
使用 tuna 命令查看 CPU 绑定
选顶: -t -P (大写) 查看特定线程的信息，例如调度策略，优先级，CPU 绑定等