Runtime options with Memory, CPUs, and GPUs
By default, a container has no resource constraints and can use as much of a
given resource as the host’s kernel scheduler allows. Docker provides ways
to control how much memory, or CPU a container can use, setting runtime
configuration flags of the
docker run command. This section provides details
on when you should set such limits and the possible implications of setting them.
Many of these features require your kernel to support Linux capabilities. To
check for support, you can use the
docker info command. If a capability
is disabled in your kernel, you may see a warning at the end of the output like
WARNING: No swap limit support
Consult your operating system’s documentation for enabling them. See also the Docker Engine troubleshooting guide for more information.
Understand the risks of running out of memory
It is important not to allow a running container to consume too much of the
host machine’s memory. On Linux hosts, if the kernel detects that there is not
enough memory to perform important system functions, it throws an
Out Of Memory Exception, and starts killing processes to free up
memory. Any process is subject to killing, including Docker and other important
applications. This can effectively bring the entire system down if the wrong
process is killed.
Docker attempts to mitigate these risks by adjusting the OOM priority on the
Docker daemon so that it is less likely to be killed than other processes
on the system. The OOM priority on containers is not adjusted. This makes it more
likely for an individual container to be killed than for the Docker daemon
or other system processes to be killed. You should not try to circumvent
these safeguards by manually setting
--oom-score-adj to an extreme negative
number on the daemon or a container, or by setting
--oom-kill-disable on a
For more information about the Linux kernel’s OOM management, see Out of Memory Management.
You can mitigate the risk of system instability due to OOME by:
- Perform tests to understand the memory requirements of your application before placing it into production.
- Ensure that your application runs only on hosts with adequate resources.
- Limit the amount of memory your container can use, as described below.
- Be mindful when configuring swap on your Docker hosts. Swap is slower and less performant than memory but can provide a buffer against running out of system memory.
- Consider converting your container to a service, and using service-level constraints and node labels to ensure that the application runs only on hosts with enough memory
Limit a container’s access to memory
Docker can enforce hard memory limits, which allow the container to use no more than a given amount of user or system memory, or soft limits, which allow the container to use as much memory as it needs unless certain conditions are met, such as when the kernel detects low memory or contention on the host machine. Some of these options have different effects when used alone or when more than one option is set.
Most of these options take a positive integer, followed by a suffix of
g, to indicate bytes, kilobytes, megabytes, or gigabytes.
||The maximum amount of memory the container can use. If you set this option, the minimum allowed value is
||The amount of memory this container is allowed to swap to disk. See
||By default, the host kernel can swap out a percentage of anonymous pages used by a container. You can set
||Allows you to specify a soft limit smaller than
||The maximum amount of kernel memory the container can use. The minimum allowed value is
||By default, if an out-of-memory (OOM) error occurs, the kernel kills processes in a container. To change this behavior, use the
For more information about cgroups and memory in general, see the documentation for Memory Resource Controller.
--memory-swap is a modifier flag that only has meaning if
--memory is also
set. Using swap allows the container to write excess memory requirements to disk
when the container has exhausted all the RAM that is available to it. There is a
performance penalty for applications that swap memory to disk often.
Its setting can have complicated effects:
--memory-swapis set to a positive integer, then both
--memory-swapmust be set.
--memory-swaprepresents the total amount of memory and swap that can be used, and
--memorycontrols the amount used by non-swap memory. So if
--memory-swap="1g", the container can use 300m of memory and 700m (
1g - 300m) swap.
--memory-swapis set to
0, the setting is ignored, and the value is treated as unset.
--memory-swapis set to the same value as
--memoryis set to a positive integer, the container does not have access to swap. See Prevent a container from using swap.
--memory-swapis unset, and
--memoryis set, the container can use as much swap as the
--memorysetting, if the host container has swap memory configured. For instance, if
--memory-swapis not set, the container can use 600m in total of memory and swap.
--memory-swapis explicitly set to
-1, the container is allowed to use unlimited swap, up to the amount available on the host system.
Inside the container, tools like
freereport the host’s available swap, not what’s available inside the container. Don’t rely on the output of
freeor similar tools to determine whether swap is present.
Prevent a container from using swap
--memory-swap are set to the same value, this prevents
containers from using any swap. This is because
--memory-swap is the amount of
combined memory and swap that can be used, while
--memory is only the amount
of physical memory that can be used.
- A value of 0 turns off anonymous page swapping.
- A value of 100 sets all anonymous pages as swappable.
- By default, if you do not set
--memory-swappiness, the value is inherited from the host machine.
Kernel memory limits are expressed in terms of the overall memory allocated to a container. Consider the following scenarios:
- Unlimited memory, unlimited kernel memory: This is the default behavior.
- Unlimited memory, limited kernel memory: This is appropriate when the amount of memory needed by all cgroups is greater than the amount of memory that actually exists on the host machine. You can configure the kernel memory to never go over what is available on the host machine, and containers which need more memory need to wait for it.
- Limited memory, unlimited kernel memory: The overall memory is limited, but the kernel memory is not.
- Limited memory, limited kernel memory: Limiting both user and kernel memory can be useful for debugging memory-related problems. If a container is using an unexpected amount of either type of memory, it runs out of memory without affecting other containers or the host machine. Within this setting, if the kernel memory limit is lower than the user memory limit, running out of kernel memory causes the container to experience an OOM error. If the kernel memory limit is higher than the user memory limit, the kernel limit does not cause the container to experience an OOM.
When you turn on any kernel memory limits, the host machine tracks “high water
mark” statistics on a per-process basis, so you can track which processes (in
this case, containers) are using excess memory. This can be seen per process
/proc/<PID>/status on the host machine.
By default, each container’s access to the host machine’s CPU cycles is unlimited. You can set various constraints to limit a given container’s access to the host machine’s CPU cycles. Most users use and configure the default CFS scheduler. You can also configure the realtime scheduler.
Configure the default CFS scheduler
The CFS is the Linux kernel CPU scheduler for normal Linux processes. Several runtime flags allow you to configure the amount of access to CPU resources your container has. When you use these settings, Docker modifies the settings for the container’s cgroup on the host machine.
||Specify how much of the available CPU resources a container can use. For instance, if the host machine has two CPUs and you set
||Specify the CPU CFS scheduler period, which is used alongside
||Impose a CPU CFS quota on the container. The number of microseconds per
||Limit the specific CPUs or cores a container can use. A comma-separated list or hyphen-separated range of CPUs a container can use, if you have more than one CPU. The first CPU is numbered 0. A valid value might be
||Set this flag to a value greater or less than the default of 1024 to increase or reduce the container’s weight, and give it access to a greater or lesser proportion of the host machine’s CPU cycles. This is only enforced when CPU cycles are constrained. When plenty of CPU cycles are available, all containers use as much CPU as they need. In that way, this is a soft limit.
If you have 1 CPU, each of the following commands guarantees the container at most 50% of the CPU every second.
$ docker run -it --cpus=".5" ubuntu /bin/bash
Which is the equivalent to manually specifying
$ docker run -it --cpu-period=100000 --cpu-quota=50000 ubuntu /bin/bash
Configure the realtime scheduler
You can configure your container to use the realtime scheduler, for tasks which cannot use the CFS scheduler. You need to make sure the host machine’s kernel is configured correctly before you can configure the Docker daemon or configure individual containers.
CPU scheduling and prioritization are advanced kernel-level features. Most users do not need to change these values from their defaults. Setting these values incorrectly can cause your host system to become unstable or unusable.
Configure the host machine’s kernel
CONFIG_RT_GROUP_SCHED is enabled in the Linux kernel by running
zcat /proc/config.gz | grep CONFIG_RT_GROUP_SCHED or by checking for the
existence of the file
/sys/fs/cgroup/cpu.rt_runtime_us. For guidance on
configuring the kernel realtime scheduler, consult the documentation for your
Configure the Docker daemon
To run containers using the realtime scheduler, run the Docker daemon with
--cpu-rt-runtime flag set to the maximum number of microseconds reserved
for realtime tasks per runtime period. For instance, with the default period of
1000000 microseconds (1 second), setting
--cpu-rt-runtime=950000 ensures that
containers using the realtime scheduler can run for 950000 microseconds for every
1000000-microsecond period, leaving at least 50000 microseconds available for
non-realtime tasks. To make this configuration permanent on systems which use
systemd, see Control and configure Docker with systemd.
Configure individual containers
You can pass several flags to control a container’s CPU priority when you
start the container using
docker run. Consult your operating system’s
documentation or the
ulimit command for information on appropriate values.
||Grants the container the
||The maximum number of microseconds the container can run at realtime priority within the Docker daemon’s realtime scheduler period. You also need the
||The maximum realtime priority allowed for the container. You also need the
The following example command sets each of these three flags on a
$ docker run -it \ --cpu-rt-runtime=950000 \ --ulimit rtprio=99 \ --cap-add=sys_nice \ debian:jessie
If the kernel or Docker daemon is not configured correctly, an error occurs.
Access an NVIDIA GPU
Visit the official NVIDIA drivers page to download and install the proper drivers. Reboot your system once you have done so.
Verify that your GPU is running and accessible.
Follow the instructions at (https://nvidia.github.io/nvidia-container-runtime/) and then run this command:
$ apt-get install nvidia-container-runtime
nvidia-container-runtime-hook is accessible from
$ which nvidia-container-runtime-hook
Restart the Docker daemon.
Expose GPUs for use
--gpus flag when you start a container to access GPU resources.
Specify how many GPUs to use. For example:
$ docker run -it --rm --gpus all ubuntu nvidia-smi
Exposes all available GPUs and returns a result akin to the following:
+-----------------------------------------------------------------------------+ | NVIDIA-SMI 384.130 Driver Version: 384.130 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GRID K520 Off | 00000000:00:03.0 Off | N/A | | N/A 36C P0 39W / 125W | 0MiB / 4036MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+
device option to specify GPUs. For example:
$ docker run -it --rm --gpus device=GPU-3a23c669-1f69-c64e-cf85-44e9b07e7a2a ubuntu nvidia-smi
Exposes that specific GPU.
$ docker run -it --rm --gpus '"device=0,2"' ubuntu nvidia-smi
Exposes the first and third GPUs.
NVIDIA GPUs can only be accessed by systems running a single engine.
Set NVIDIA capabilities
You can set capabilities manually. For example, on Ubuntu you can run the following:
$ docker run --gpus 'all,capabilities=utility' --rm ubuntu nvidia-smi
This enables the
utility driver capability which adds the
nvidia-smi tool to
Capabilities as well as other configurations can be set in images via environment variables. More information on valid variables can be found at the nvidia-container-runtime GitHub page. These variables can be set in a Dockerfile.
You can also utitize CUDA images which sets these variables automatically. See the CUDA images GitHub page for more information.