cgroups: CPU Limits

CPU limits in cgroups come in three flavors:

Shares (relative weight):

cpu.shares = 1024  (default)

Two cgroups with shares 1024 each split CPU 50/50 under contention. With shares 2048 vs 1024, it's 67/33. No effect when only one is using CPU.

Quota (hard cap on absolute CPU time):

cpu.cfs_period_us = 100000     (100ms accounting period)
cpu.cfs_quota_us  = 50000      (50ms allowed per period)

Cap at 0.5 CPUs (50%). Period is fixed at 100ms by default; quota is the work budget. Setting quota > period * num_cores means "more than one core's worth allowed".

Cpuset (specific CPUs):

cpuset.cpus = 0-3        (this cgroup runs only on cores 0-3)
cpuset.mems = 0          (NUMA node 0 only)

Used for cache-locality-sensitive workloads. Pin a real-time process to its own cores.

In Docker:

• --cpu-shares 512 -> cpu.shares=512
• --cpus 0.5 -> cfs_quota=50000, cfs_period=100000
• --cpuset-cpus 0,2 -> cpuset.cpus=0,2

Throttling is invisible to the process — it just runs slower than expected. Watch throttled_time in cpu.stat for evidence.

For this lesson: model quota-based CPU limiting.