DEPRECATION NOTICE:
These scripts are useless on a modern linux with unified cgroup-v2 hierarchy.
On a systemd-enabled machine, use "systemd-run --scope" instead.
See also: mk-fg/fgtk#cgrc tool
A set of tools to work with cgroup tree and process classification/QoS according to it.
More (of a bit outdated) info can be found in a blog post here.
Main script there - cgconf - allows to use YAML like this to configure initial cgroup hierarcy like this:
path: /sys/fs/cgroup
defaults:
_tasks: root:wheel:664
_admin: root:wheel:644
# _path: root:root:755 # won't be chown/chmod'ed, if unset
freezer:
groups:
# Must be applied to root cgroup
memory.use_hierarchy: 1
# base:
# _default: true # to put all pids here initially
# cpu.shares: 1000
# blkio.weight: 1000
user.slice:
cpu:
cfs_quota_us: 1_500_000
cfs_period_us: 1_000_000
tagged:
cave:
_tasks: root:paludisbuild
_admin: root:paludisbuild
cpu:
shares: 100
cfs_quota_us: 100_000
cfs_period_us: 250_000
blkio.weight: 100
memory.soft_limit_in_bytes: 2G
desktop:
roam:
_tasks: root:users
cpu.shares: 300
blkio.weight: 300
memory.soft_limit_in_bytes: 2G
de_misc:
memory.soft_limit_in_bytes: 700M
memory.limit_in_bytes: 1500M
vm:
quasi:
misc:
cpu.shares: 200
blkio.weight: 100
memory.soft_limit_in_bytes: 1200M
memory.limit_in_bytes: 1500M
bench:
# Subdir for adhoc cgroups created by user
tmp:
# Corresponding pw_gid will be used, if "user:" is specified
# Specs like "user", ":group:770" or "::775" are all valid.
_tasks: 'fraggod:'
_admin: 'fraggod:'
_path: 'fraggod:'
# These will be initialized as dirs with proper uid/gid, but no stuff applied there
cpuacct:
memory:
blkio:
# Limits that groups in tmp/ can't transcend
cpu.shares: 500
blkio.weight: 500
memory.soft_limit_in_bytes: 300M
memory.limit_in_bytes: 500MAnd then something like cgrc <tagged-name> <cmd> <args...> to run anything inside these (can also be used in shebang with -s, with cmd from stdin, etc).
Other tools allow waiting for threads within some cgroup to finish before proceeding (cgwait), put stuff running there on hold easily (cgfreeze) and run stuff in temp-cgroup, reporting accounting data for it afterwards (cgtime).
cgconf and cgrc turn out to be surprisingly useful still, despite systemd adding knobs to control cgroup resource limits (but not all of them, and spread over lot of small files, which are pain if you need a big picture of e.g. weights) and systemd-run, which hides i/o of whatever it runs in systemd slices.