Add cpu-shares to docker jobs

LSF sets for every job the cpu.weight in cgroup v2 to make sure each job uses the CPUs based on number of requested tasks and total slots on host running the job. This is not the case with docker jobs. All processes created by docker have default cpu.weight of 100 which means that the docker can use more CPU than it should. This can create a load on the host and also does not keep the original CPU weight the job has in the current host. The docker flag cpu-shares should be passed in the "docker run" command that docker-starter.py script creates, same as it passes the memory limit to the docker.

Needed By

Yesterday (Let's go already!)

Post comment

Admin

Bill McMillan

Apr 30, 2026

For cgroupv1 docker defaults to cgroupfs as the cgroup driver. When cgroupv2 is enabled, or systemd is running, it will default to systemd being the cgroup driver. This means that LSF cannot apply controls as it is not the cgroup owner.

You can change the docker config such that cgroupfs will be used, even if systemd is running.

Modify (or create if it does not exist) /etc/docker/daemon.json with the following contents:
{
"exec-opts": ["native.cgroupdriver=cgroupfs"]
}

Restart docker:
# systemctl restart docker

Verify that cgroupfs is being used as the cgorup driver:
# docker info | grep -i 'cgroup driver'
Cgroup Driver: cgroupfs

Submit a docker job to this host and verify that the cpu.weight is set based on "bsub -n":
/sys/fs/cgroup/lsf/<cluster_name>/job.xxx.xxxx/

Reply
Hide replies

By clicking the "Post Comment" or "Submit Idea" button, you are agreeing to the IBM Ideas Portal Terms of Use.
Do not place IBM confidential, company confidential, or personal information into any field.

Please enter your email address

RELATED IDEAS

Add cpu-shares to docker jobs