Linux – Why is cpu.cfs_quota_us not limiting CPU bandwidth of LXC container

cgroupslinuxlxc

I'd like to limit the container to 25% of the system's total CPU bandwidth.
Here's my setup:

  • LXC version 1.0.2
  • kernel 3.2.45
  • one user created cgroup (foo) for an LXC container
  • 40 available cores on the host
  • the host and container have default values for every other cgroup subsystem except:
  • /sys/fs/cgroup/cpu/lxc/foo/cpu.cfs_quota_us = 400000
  • /sys/fs/cgroup/cpu/lxc/foo/cpu.cfs_period_us = 100000
  • /sys/fs/cgroup/cpuset/lxc/foo/cpuset.cpus = 0-15

I calculated the quota using this formula:

(# of cpus available to container) * (cpu.cfs_period_us) * (.25) so 16 * 100000 * .25 = 400000

I ran a basic stress-ng inside and outside the container at the same time to get a gauge of how many operations per second were being allowed inside and out and the results were basically the same as running with a quota of "-1", which is to say no quota.

Outside Run:

$ ./stress-ng  --cpu-load 50 -c 40 --timeout 20s --metrics-brief
stress-ng: info: [25649] dispatching hogs: 40 cpu  
stress-ng: info: [25649] successful run completed in 20.44s  
stress-ng: info: [25649] stressor      bogo ops real time  usr time  sys time   bogo ops/s   bogo ops/s  
stress-ng: info: [25649]                          (secs)    (secs)    (secs)   (real time) (usr+sys time)  
stress-ng: info: [25649] cpu              37348     20.18    380.56      0.58      1850.85        97.99

Inside Run:

$ ./stress-ng --cpu-load 100 -c 16 --timeout 20s --metrics-brief  
stress-ng: info: [34256] dispatching hogs: 16 cpu  
stress-ng: info: [34256] successful run completed in 20.10s  
stress-ng: info: [34256] stressor      bogo ops real time  usr time  sys time   bogo ops/s   bogo ops/s  
stress-ng: info: [34256]                          (secs)    (secs)    (secs)   (real time) (usr+sys time)  
stress-ng: info: [34256] cpu              24147     20.03    205.20      0.17      1205.67       117.58

Based on the ops/s I'm getting 39%. Why does this happen? Shouldn't it be limited by cpu.cfs_quota_us?

Thanks for the help in advance.

Best Answer

Wanted to post the answer to this question in case anyone else sees a similar confusing result. It looks like I had two problems:

  • Need to use # of cpus on host, not # available CPUs in the cgroups cpuset to estimate CPU bandwidth:

    (# of cpus on the host) * (cpu.cfs_period_us) * (.25) so 40 * 100000 * .25 = 1000000

  • My run of stress-ng inside the container was using the cpu and cpuset controllers of the /lxc/foo cgroup while the run of stress-ng outside of the container was using the /system/sshd.service cgroup

To better model my real world application I should have specified which controllers to use by using cgexec: 

$ cgexec -g cpuset:/lxc/foo -g cpu:/lxc/foo ./stress-ng --cpu-load 100 -c 48 --times --timeout 10s --metrics-brief  
stress-ng: info: [6252] dispatching hogs: 48 cpu  
stress-ng: info: [6252] successful run completed in 10.36s  
stress-ng: info: [6252] stressor      bogo ops real time  usr time  sys time   bogo ops/s   bogo ops/s  
stress-ng: info: [6252]                          (secs)    (secs)    (secs)   (real time) (usr+sys time)  
stress-ng: info: [6252] cpu              11152     10.09    102.83      0.12      1105.60       108.32  
stress-ng: info: [6252] for a 10.36s run time:  
stress-ng: info: [6252]     414.46s available CPU time  
stress-ng: info: [6252]     102.85s user time   ( 24.82%)  
stress-ng: info: [6252]       0.12s system time (  0.03%)  
stress-ng: info: [6252]     102.97s total time  ( 24.84%)
Related Question