Skip to content

Latest commit

 

History

History
98 lines (76 loc) · 3.03 KB

02-cgroups-cpu.md

File metadata and controls

98 lines (76 loc) · 3.03 KB

Cgroups

The goal of this hands-on session is to get familiar with configuring cgroups to control resource allocation to processes. We will use the cpu cgroup as the target resource.

In particular, you will learn how to adjust the proportion of CPU cycles that a process gets over another.

NOTE: the scenario below is valid if the platform you are working on has only 1 CPU. If your platform has more than 1 CPUs, then you will need in addition to restrict the processes so that they are scheduled one the same CPU, e.g. CPU0. In that case, follow the instructions under backup/cgroups.md.

  1. At first, inspect the contents of the cpu group

    ls /sys/fs/cgroup/cpu/

    This is the "root" group for the cpu subsystem, meaning that every process initially belongs there, since no special adjustments have been yet made to change the portion of CPU time that processes get. The processes belonging to a cgroup are registered in the tasks file, so initially the following file should contain all the PIDs of the system:

    cat /sys/fs/cgroup/cpu/tasks

    Under the cpu subsystem, you can see a number of tunable parameters, like cpu.shares. cpu.shares specifies a relative share of CPU time available to the tasks in a cgroup.

    What is the current value of this parameter in the root group?

    cat /sys/fs/cgroup/cpu/cpu.shares

  2. Create two control groups under the cpu subsystem, a "realtime" and a "batch", making all the tunable parameters of the cpu subsystem writable by current user

    sudo cgcreate -a $USER -g cpu:realtime
sudo cgcreate -a $USER -g cpu:batch

    Notice that two new subdirs have been created under the root cpu hierarchy:

    ls /sys/fs/cgroup/cpu/

    As no special adjustments have been yet made, the default values for cpu.shares should be those of the parent (i.e. "root"). Verify it:

    cat /sys/fs/cgroup/cpu/realtime/cpu.shares
cat /sys/fs/cgroup/cpu/batch/cpu.shares

  3. Launch two CPU intensive processes

    dd if=/dev/zero of=/dev/null &
dd if=/dev/zero of=/dev/null &

and get their PIDs pgrep -x dd

  1. Move each process to a separate cgroup

    sudo cgclassify -g cpu:realtime <PID1>
sudo cgclassify -g cpu:batch <PID2>

    What is their CPU usage?

    top

  2. Change their CPU shares so that the process in the "realtime" group gets 75% of CPU 0 time, while the process in the "batch" group gets 25% of CPU 0 time.

    NOTE: it is the relative weights of cpu.shares that matter, not their absolute values.

    HINT: use values larger than 10

    echo ??? > /sys/fs/cgroup/cpu/realtime/cpu.shares 
echo ??? > /sys/fs/cgroup/cpu/batch/cpu.shares

    What is their CPU usage now?

    top

  3. Stop processes and delete cgroups

    killall dd 
sudo cgdelete -g cpu:realtime -g cpu:batch