Hi, I recently found similar behavior in an app for our company. A simple threaded cpu benchmark shows:

% numactl -C 0,5 ./ssp 12 elapsed time: 99943 ms

cpu.cfs_quota_us = 200000 cpu.cfs_period_us = 100000 % cgexec -g cpu:cgtestq ./ssp 12 elapsed time: 420888 ms

cpu.cfs_quota_us = 2000 cpu.cfs_period_us = 1000 % cgexec -g cpu:cgtestqx ./ssp 12 elapsed time: 168104 ms

Also interesting was in our app some RR thread priorities are used, and those do not get controlled via the cgroup cpu.cfs settings.