Limiting Container Resources
It’s often useful to limit the resources that are consumed by a container, e.g. to allow the container to only use 1 CPU, or 50% of the RAM on the system. Although on HPC clusters it’s common to launch containers with a job scheduler that can set limits per job, the following scenarios are examples where more direct control is useful:
When running multiple containerized applications inside an HPC job, where each container in the job should have different resource limits.
When testing HPC code on a workstation, to avoid excessive CPU / RAM usage bringing the desktop environment and other applications to a halt.
When benchmarking code that doesn’t provide internal means to limit the number of CPUs it uses.
There are three ways to apply limits to a container that is run with SingularityCE:
Using the command line flags introduced in v3.10.
--apply-cgroupsflag to apply a
cgroups.tomlfile that defines the resource limits.
Using external tools such as
systemd-runtool to apply limits, and then call
Requirements - Linux Cgroups
Resource limits are applied to containers using functionality in the Linux kernel known as control groups or cgroups. There are two versions of cgroups:
cgroups v1 has a more complex structure, and allows only the root user to
safely apply limits to applications. If your system is using cgroups v1 then you
can only use the CLI resource limit flags or
--apply-cgroups when running
containers as the root user.
cgroups v2 has a simplified structure, and is designed in a way that permits delegation of cgroups control to standard users. This delegation is usually accomplished via systemd.
Generally, to apply resource limits to a container as a non-root user your system must:
Be using cgroups v2, in the unified hierarchy mode.
Have a Linux kernel version >= 4.15.
Have systemd version >= 224.
systemd cgroupsenabled in
singularity.conf(this is the default).
Have systemd configured to delegate cgroups controllers to non-root users.
Recent distributions such as Ubuntu 22.04, Debian 11, Fedora 31, and newer, satisfy these criteria by default. On older distributions support can often be enabled. Consult the admin documentation or speak to your system administrator about this.
Command Line Limit Flags
SingularityCE 3.10 introduced a number of simple command line flags that you can use with shell/run/exec and the instance commands to directly apply resource limits to a container when you run it.
The flags detailed below are compatible with those used by the
except that the short forms are not supported. For example, you cannot use
-c instead of
-c is used by SingularityCE for
Not all limits provided by other runtimes are currently supported by
SingularityCE. Specifically, the
--device- flags supported by the
CLI are not yet available.
--cpus sets the number of CPUs, or fractional CPUs, that the container can
use. The minimum is
0.01 or one hundredth of a physical CPU. The maximum is
the number of CPU cores on your system.
# Limit container to 3.5 CPUs $ singularity run --cpus 3.5 myfirstapp.sif
--cpu-shares sets a relative weight for a container’s access to the system’s
CPUs, versus other containers that also have a
--cpus-shares value set. If
container A has 1024 cpu shares, and container B has 512 cpu shares, then
container A will receive twice as much CPU time than container B, but only when
there is contention for CPUs, i.e. the containers are able to consume more CPU
time than is available.
# Container A - twice as much CPU priority as container B $ singularity run --cpu-shares 1024 myfirstapp.sif # Container A - half as much CPU priority as container A $ singularity run --cpu-shares 512 mysecondapp.sif
--cpu-set-cpus specifies a list of physical CPU IDs on which a container can
run. For example, on a dual CPU system you might pin one container to the first
12 cores on CPU 1, and a second container to the second 12 cores on CPU 2.
--cpu-set-mems specifies a list of memory nodes the container can use. It
should generally be set to the same value as
# Container A - first CPU $ singularity run --cpu-set-cpus 0-11 --cpu-set-mems 0-11 myfirstapp.sif # Container B - second CPU $ singularity run --cpu-set-cpus 12-23 --cpu-set-mems 12-23 myfirstapp.sif
--memory sets the maximum amount of RAM that a container can use, in bytes.
You can use suffixes such as
G to specify megabytes or gigabytes.
If the container tries to use more memory than its limit, the system will kill
# Run a program that will use 10GB of RAM, with a 100MB limit $ singularity exec --memory 100M memhog.sif memhog 10G .................................................Killed
--memory-reservation sets a soft limit, which should be lower than the hard
limit set with
--memory. When there is contention for memory, the system
will attempt to make sure the container doesn’t exceed the soft limit.
# Kill my program if it exceeds 10G, but aim for 8G if there is contention $ singularity exec --memory 10G --memory-reservation 8G myfirstapp.sif
--memory-swap sets the total amount of memory and swap space that a
container can use. You must set
--memory along with
-1 means unlimited swap. If
--memory-swap is not set or is
0, then the container can use an amount of swap up to the value of
--memory. It’s easier to understand this flag with examples:
# Run a container that can use up to 1G RAM, or swap if it is swapped out $ singularity run --memory 1G myfirstapp.sif # Run a container that can use up to 1G RAM, and no swap space $ singularity run --memory 1G --memory-swap 1G myfirstapp.sif # Run a container that can use up to 1G RAM, and unlimited swap space $ singularity run --memory 1G --memory-swap -1 myfirstapp.sif # Run a container that can use up to 1G RAM, and 1G swap space $ singularity run --memory 1G --memory-swap 2G myfirstapp.sif
bfq IO scheduler to be configured for block IO on
the system. This is common on modern distributions, but not universal. Ask
your system administrator if IO limits are not working as expected.
--blkio-weight sets a relative weight for the container when performing
block I/O, e.g. reading/writing to/from disk. The weight should be between 10
and 1000, and will control how much I/O access a container recieves when there
is contention for I/O with other containers. It may be useful to give high
priority to a container that needs infrequent but time sensitive data access,
running alongside an application that is continuously performing bulk reads.
# Container A - ten times as much block IO priority as container B $ singularity run --blkio-weight 1000 myfirstapp.sif # Container A - ten times less block IO priority as container A $ singularity run --blkio-weight 100 mysecondapp.sif
--blkio-weight-device sets a relative weight for the container when performing
block I/O on a specific device. Specify the device and weight as
# Container A - ten times as much block IO priority as container B on disk /dev/sda $ singularity run --blkio-weight-device /dev/sda:1000 myfirstapp.sif # Container A - ten times less block IO priority as container A on disk /dev/sda $ singularity run --blkio-weight-device /dev/sda:100 mysecondapp.sif
Applying Resource Limits From a TOML file
SingularityCE 3.9 and above can directly apply resource limitations to systems
configured for both cgroups v1 and the v2 unified hierarchy, using the
--apply-cgroups flag. Resource limits are specified using a TOML file that
represents the resources section of the OCI runtime-spec:
On a cgroups v1 system the resources configuration is applied directly. On a cgroups v2 system the configuration is translated and applied to the unified hierarchy.
Under cgroups v1, access restrictions for device nodes are managed directly. Under cgroups v2, the restrictions are applied by attaching eBPF programs that implement the requested access controls.
To apply resource limits to a container, using the
flag, which takes a path to a TOML file specifying the cgroups
configuration to be applied:
$ singularity shell --apply-cgroups /path/to/cgroups.toml my_container.sif
--apply-cgroups as a non-root user requires a cgroups v2 system,
configured to use the
systemd cgroups manager in
CPU usage can be limited using different strategies, with limits
specified in the
[cpu] section of the TOML file.
This corresponds to a ratio versus other cgroups with cpu shares.
Usually the default value is
1024. That means if you want to allow
to use 50% of a single CPU, you will set
512 as value.
[cpu] shares = 512
A cgroup can get more than its share of CPU if there are enough idle CPU cycles available in the system, due to the work conserving nature of the scheduler, so a contained process can consume all CPU cycles even with a ratio of 50%. The ratio is only applied when two or more processes conflicts with their needs of CPU cycles.
You can enforce hard limits on the CPU cycles a cgroup can consume, so
contained processes can’t use more than the amount of CPU time set for
quota allows you to configure the amount of CPU time
that a cgroup can use per period. The default is 100ms (100000us). So if
you want to limit amount of CPU time to 20ms during period of 100ms:
[cpu] period = 100000 quota = 20000
You can also restrict access to specific CPUs (cores) and associated
memory nodes by using
[cpu] cpus = "0-1" mems = "0-1"
Where the container has limited access to CPU 0 and CPU 1.
It’s important to set identical values for both
To limit the amount of memory that your container uses to 500MB
(524288000 bytes), set a
limit value inside the
of your cgroups TOML file:
[memory] limit = 524288000
Start your container, applying the toml file, e.g.:
$ singularity run --apply-cgroups path/to/cgroups.toml library://alpine
After that, you can verify that the container is only using 500MB of
memory. This example assumes that there is only one running container.
If you are running multiple containers you will find multiple cgroups
trees under the
# cgroups v1 $ cat /sys/fs/cgroup/memory/singularity/*/memory.limit_in_bytes 524288000 # cgroups v2 - note translation of memory.limit_in_bytes -> memory.max $ cat /sys/fs/cgroup/singularity/*/memory.max 524288000
To control block device I/O, applying limits to competing container, use
[blockIO] section of the TOML file:
[blockIO] weight = 1000 leafWeight = 1000
leafWeight accept values between
weight is the default weight of the group on all the devices until
and unless overridden by a per device rule.
leafWeight relates to weight for the purpose of deciding how heavily
to weigh tasks in the given cgroup while competing with the cgroup’s
To apply limits to specific block devices, you must set configuration
for specific device major/minor numbers. For example, to override
devices, set limits for device major 7, minor 0 and 1:
[blockIO] [[blockIO.weightDevice]] major = 7 minor = 0 weight = 100 leafWeight = 50 [[blockIO.weightDevice]] major = 7 minor = 1 weight = 100 leafWeight = 50
You can also limit the IO read/write rate to a specific absolute value,
e.g. 16MB per second for the
/dev/loop0 block device. The
is specified in bytes per second.
[blockIO] [[blockIO.throttleReadBpsDevice]] major = 7 minor = 0 rate = 16777216 [[blockIO.throttleWriteBpsDevice]] major = 7 minor = 0 rate = 16777216
Device limits can only be applied when running as the root user, and will be ignored as a non-root user.
You can limit read (
r), write (
w), or creation (
devices by a container. Like applying I/O limits to devices, you must
use device node major and minor numbers to create rules for specific
devices or classes of device.
In this example, a container is configured to only be able to read from
or write to
[[devices]] access = "rwm" allow = false [[devices]] access = "rw" allow = true major = 1 minor = 3 type = "c"
SingularityCE can apply all resource limits that are valid in the OCI
resources section. If you use cgroups v1 limits on a cgroups v2
system they will be translated at runtime. You may also specify native cgroups
v2 limits under the
See https://github.com/opencontainers/runtime-spec/blob/master/config-linux.md#control-groups for information about the available limits. Note that SingularityCE uses TOML format for the configuration file, rather than JSON.
Applying Resource Limits With External Tools
Because SingularityCE starts a container as a simple process, rather
than using a daemon, you can limit resource usage by running the
singularity command inside an existing cgroup. This is convenient
where, for example, a job scheduler uses cgroups to control job limits.
singularity inside your batch script, your container will
respect the limits set by the scheduler on the job’s cgroup.
As well as schedulers you can use tools such as
create a cgroup, and run SingularityCE inside of it. This is convenient
on modern cgroups v2 systems, where the creation of cgroups can be
delegated to users through systemd. Without this delegation
privileges are required to create a cgroup.
For example, assuming your system is configured correctly for unprivileged cgroup creation via systemd, you can limit the number of CPUs a container run is allowed to use:
$ systemd-run --user --scope -p AllowedCPUs=1,2 -- singularity run mycontainer.sif
--userinstructs systemd that we want to run as our own user account.
--scopewill run our command in an interactive scope that inherits from our environment. By default the command would run as a service, in the background.
-p AllowedCPUs=1,2sets a property on our scope, so that in this case systemd will then setup a cgroup limiting our command to using CPU 1 and 2 only.
The double hyphen
--separates options for
systemd-runfrom the actual command we wish to execute. This is important so that
systemd-rundoesn’t capture any flags we might need to pass to
You can read more about how systemd can control resources uses at the
link below, which details the properties you can set using