Terminal cgroup, short for tcg, is a tool to help people manage system's resource usage through cgroup. To use tcg, you should first do the following:
- Prepare your system to use cgroup v2 unified hierarchy [optional, but highly recommended]
- Install
tcg. - Add something like
printf "Your cgroup is "; tcg createto your.bashrcor.zshrc, so that it runs everytime you open a new shell - Read the example use case below.
Starting from systemd v248, systemd now defaults to the "unified" cgroup hierarchy.
For systems prior to v248, add the following kernel parameters at boot time:
systemd.unified_cgroup_hierarchy=1- Systemd will mount/sys/fs/cgroupas cgroup v2cgroup_no_v1="all"- The kernel will disable all v1 cgroup controllers
If you use Archlinux, you can install from Archlinux AUR.
If you use Ubuntu or Debian, you can add the following line to your /etc/apt/sources.list:
deb [trusted=yes check-date=no] https://zasdfgbnm.github.io/tcg/deb/ all main
and install with apt install tcg.
If you use Xonsh, we have a xontrib, see shells/xonsh/README.md for more information.
tcg requires boost, fmt, and spdlog, you need to install them on your system first.
Also, tcg is written with C++20, currently only gcc >=10.2 supports all the features that tcg uses.
To clone tcg from github, run the following command:
git clone https://github.com/zasdfgbnm/tcg.git
After you cloned tcg, go to tcg's directory and simply do
mkdir build
cd build
cmake ..
make -jYou will get a binary called tcg in the build directory.
The last step, you need to make sure tcg is owned by root and has suid permission:
sudo chown root tcg
sudo chmod +s tcg
In my daily job, I often need to compile PyTorch, which will take 100% of my CPU for a long time.
Since I added tcg create to my .xonshrc, everytime when I open a shell, I will get a cgroup containing the shell together with all the commands I run in that shell. The name will be displayed on the shell:
gaoxiang@XiangdeMacBook-Pro ~ $ ssh xxxxxxxx
Last login: Mon Dec 28 19:05:50 2020 from 2600:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx
Your cgroup is poincare.
gaoxiang@sunnyvale ~ $
The above example is a shell from ssh, and is assigned a cgroup named poincare.
I start to compile PyTorch in the shell, and the CPU usage of my system easily reaches 100%. Before that compilation job finishes, I decided that I want to benchmark something. To be accurate on my benchmark, I want to pause my compilation for a few minutes. To do so, I can easily do:
tcg freeze poincare
And when I am ready to put my compilation back to work, I just need to:
tcg unfreeze poincare
and the compilation will continue its work again.
Letting the compilation job to use 100% of my CPU is not always a good idea. Because this sometimes makes my computer's response time slow. To solve this problem, I can simply do
tcg set poincare cpu.max "550000 10000"
to limit the CPU usage to about 90% on my computer. Please note that the two numbers depend on the number of cores in your computer, so don't just copy and paste the number above! The above 550000 10000 means, for every 10000 periods, the cgroup are allowed to use at most 550000 periods. My computer has 32 cores and 64 threads, so it can use at most 640000 for every 10000 periods. Limiting it to 550000 will allow 90000 spare periods for other tasks.
To relax this limit, simply do:
tcg set poincare cpu.max max
this will allow the compilation job to run at 100% CPU usage
For detailed information, please check Linux kernel's cgroup v2 documentation.
Let's say I have two compilation jobs, one at cgroup poincare, another at cgroup pauli. I want these two jobs togeter use as much CPU as possible, but I think poincare is 10x more important than pauli. To let poincare get 10x more CPU time than pauli, I can do:
tcg set pauli cpu.weight 100
tcg set poincare cpu.weight 1000
or even simpler:
tcg set poincare cpu.weight 1000
The tcg set pauli cpu.weight 100 can be omitted because the default cpu.weight value for cgroups are 100.
Setting the cpu.weight value tells Linux kernel to control the number of CPU cycles taskes inside each cgroup when scheduling.
For detailed information, please check Linux kernel's cgroup v2 documentation.