When running mixed workloads—latency-sensitive services alongside background batch jobs—you need a way to deprioritize the batch cgroup so it consumes CPU only when no higher-priority peers need it. Alibaba Cloud Linux 3 provides the cpu.idle interface, the cgroup-level analog of the per-task SCHED_IDLE scheduling policy. Set cpu.idle to 1 on a cgroup to mark it as idle-priority: the cgroup competes for CPU only when no non-idle peers are runnable. This feature requires kernel version 5.10.134-16.al8 or later.
Interface file location
The cpu.idle interface file location depends on the cgroup version:
| cgroup version | Path |
|---|---|
| cgroup v1 | /sys/fs/cgroup/cpu/$cg |
| cgroup v2 | /sys/fs/cgroup/$cg |
$cg is a placeholder for the cgroup name.
Enable or disable SCHED_IDLE for a cgroup
| Interface | Value | Effect |
|---|---|---|
cpu.idle |
0 |
Disables the SCHED_IDLE scheduling policy for the cgroup. |
1 |
Enables the SCHED_IDLE scheduling policy for the cgroup. The cgroup receives the minimum time slice weight and is no longer guaranteed a minimum runtime. |
Limitations
Enabling SCHED_IDLE (setting cpu.idle to 1) has the following side effects due to incompatibilities between cgroup versions and the SCHED_IDLE policy:
| cgroup version | Limitation |
|---|---|
| cgroup v1 | The cpu.shares interface cannot be overridden. |
| cgroup v2 | The cpu.weight and cpu.weight.nice interfaces cannot be overridden. |