Amazon ECS task definitions for GPU workloads - Amazon Elastic Container Service
ConsiderationsShare GPUsWhat to do if you need a P2 instance

Amazon ECS task definitions for GPU workloads

Amazon ECS supports workloads that use GPUs, when you create clusters with container instances that support GPUs. Amazon EC2 GPU-based container instances that use the p2, p3, p5, g3, g4, and g5 instance types provide access to NVIDIA GPUs. For more information, see Linux Accelerated Computing Instances in the Amazon EC2 Instance Types guide.

Amazon ECS provides a GPU-optimized AMI that comes with pre-configured NVIDIA kernel drivers and a Docker GPU runtime. For more information, see Amazon ECS-optimized Linux AMIs.

You can designate a number of GPUs in your task definition for task placement consideration at a container level. Amazon ECS schedules to available container instances that support GPUs and pin physical GPUs to proper containers for optimal performance.

The following Amazon EC2 GPU-based instance types are supported. For more information, see Amazon EC2 P2 Instances, Amazon EC2 P3 Instances, Amazon EC2 P4d Instances, Amazon EC2 P5 Instances, Amazon EC2 G3 Instances, Amazon EC2 G4 Instances, Amazon EC2 G5 Instances and Amazon EC2 G6 Instances.

Instance type GPUs GPU memory (GiB) vCPUs Memory (GiB)

p3.2xlarge

1

16

8

61

p3.8xlarge

4

64

32

244

p3.16xlarge

8

128

64

488

p3dn.24xlarge

8

256

96

768

p4d.24xlarge

 8 320 96 1152
p5.48xlarge 8 640 192 2048

g3s.xlarge

1

8

4

30.5

g3.4xlarge

1

8

16

122

g3.8xlarge

2

16

32

244

g3.16xlarge

4

32

64

488

g4dn.xlarge

1

16

4

16

g4dn.2xlarge

1

16

8

32

g4dn.4xlarge

1

16

16

64

g4dn.8xlarge

1

16

32

128

g4dn.12xlarge

4

64

48

192

g4dn.16xlarge

1

16

64

256

g5.xlarge

1

24

4

16

g5.2xlarge

1

24

8

32

g5.4xlarge

1

24

16

64

g5.8xlarge

1

24

32

128

g5.16xlarge

1

24

64

256

g5.12xlarge

4

96

48

192

g5.24xlarge

4

96

96

384

g5.48xlarge

8

192

192

768
g6.xlarge 1 24 4 16
g6.2xlarge 1 24 8 32
g6.4xlarge 1 24 16 64
g6.8xlarge 1 24 32 128
g6.16.xlarge 1 24 64 256
g6.12xlarge 4 96 48 192
g6.24xlarge 4 96 48 192
g6.48xlarge 8 192 192 768
g6.metal 8 192 192 768
gr6.4xlarge 1 24 16 128
gr6.8xlarge 1 24 32 256

You can retrieve the Amazon Machine Image (AMI) ID for Amazon ECS-optimized AMIs by querying the AWS Systems Manager Parameter Store API. Using this parameter, you don't need to manually look up Amazon ECS-optimized AMI IDs. For more information about the Systems Manager Parameter Store API, see GetParameter. The user that you use must have the ssm:GetParameter IAM permission to retrieve the Amazon ECS-optimized AMI metadata.

aws ssm get-parameters --names /aws/service/ecs/optimized-ami/amazon-linux-2/gpu/recommended --region us-east-1

Considerations

Note

The support for g2 instance family type has been deprecated.

The p2 instance family type is only supported on versions earlier than 20230912 of the Amazon ECS GPU-optimized AMI. If you need to continue to use p2 instances, see What to do if you need a P2 instance.

In-place updates of the NVIDIA/CUDA drivers on both these instance family types will cause potential GPU workload failures.

We recommend that you consider the following before you begin working with GPUs on Amazon ECS.

  • Your clusters can contain a mix of GPU and non-GPU container instances.

  • You can run GPU workloads on external instances. When registering an external instance with your cluster, ensure the --enable-gpu flag is included on the installation script. For more information, see Registering an external instance to an Amazon ECS cluster.

  • You must set ECS_ENABLE_GPU_SUPPORT to true in your agent configuration file. For more information, see Amazon ECS container agent configuration.

  • When running a task or creating a service, you can use instance type attributes when you configure task placement constraints to determine the container instances the task is to be launched on. By doing this, you can more effectively use your resources. For more information, see How Amazon ECS places tasks on container instances.

    The following example launches a task on a g4dn.xlarge container instance in your default cluster.

    aws ecs run-task --cluster default --task-definition ecs-gpu-task-def \
     --placement-constraints type=memberOf,expression="attribute:ecs.instance-type ==  g4dn.xlarge" --region us-east-2

  • For each container that has a GPU resource requirement that's specified in the container definition, Amazon ECS sets the container runtime to be the NVIDIA container runtime.

  • The NVIDIA container runtime requires some environment variables to be set in the container to function properly. For a list of these environment variables, see Specialized Configurations with Docker. Amazon ECS sets the NVIDIA_VISIBLE_DEVICES environment variable value to be a list of the GPU device IDs that Amazon ECS assigns to the container. For the other required environment variables, Amazon ECS doesn't set them. So, make sure that your container image sets them or they're set in the container definition.

  • The p5 instance type family is supported on version 20230929 and later of the Amazon ECS GPU-optimized AMI.

  • The g4 instance type family is supported on version 20230913 and later of the Amazon ECS GPU-optimized AMI. For more information, see Amazon ECS-optimized Linux AMIs. It's not supported in the Create Cluster workflow in the Amazon ECS console. To use these instance types, you must either use the Amazon EC2 console, AWS CLI, or API and manually register the instances to your cluster.

  • The p4d.24xlarge instance type only works with CUDA 11 or later.

  • The Amazon ECS GPU-optimized AMI has IPv6 enabled, which causes issues when using yum. This can be resolved by configuring yum to use IPv4 with the following command.

    echo "ip_resolve=4" >> /etc/yum.conf

  • When you build a container image that doesn't use the NVIDIA/CUDA base images, you must set the NVIDIA_DRIVER_CAPABILITIES container runtime variable to one of the following values:

    • utility,compute

    • all

    For information about how to set the variable, see Controlling the NVIDIA Container Runtime on the NVIDIA website.

  • GPUs are not supported on Windows containers.

Share GPUs

Add the following user data to your instances when you want multiple container to share 1 GPU. . For more information see Run commands when you launch an EC2 instance with user data input in the Amazon EC2 User Guide.

Use the last supported GPU-optimized AMI

You can use the 20230906 version of the GPU-optimized AMI, and add the following to the instance user data.

Replace cluster-name with the name of your cluster.

const userData = ec2.UserData.forLinux();
 userData.addCommands(
 'sudo rm /etc/sysconfig/docker',
 'echo DAEMON_MAXFILES=1048576 | sudo tee -a /etc/sysconfig/docker',
 'echo OPTIONS="--default-ulimit nofile=32768:65536 --default-runtime nvidia" | sudo tee -a /etc/sysconfig/docker',
 'echo DAEMON_PIDFILE_TIMEOUT=10 | sudo tee -a /etc/sysconfig/docker',
 'sudo systemctl restart docker',
);

What to do if you need a P2 instance

If you need to use P2 instance, you can use one of the following options to continue using the instances.

You must modify the instance user data for both options. For more information see Run commands when you launch an EC2 instance with user data input in the Amazon EC2 User Guide.

Use the last supported GPU-optimized AMI

You can use the 20230906 version of the GPU-optimized AMI, and add the following to the instance user data.

Replace cluster-name with the name of your cluster.

#!/bin/bash
echo "exclude=*nvidia* *cuda*" >> /etc/yum.conf
echo "ECS_CLUSTER=cluster-name" >> /etc/ecs/ecs.config

Use the latest GPU-optimized AMI, and update the user data

You can add the following to the instance user data. This uninstalls the Nvidia 535/Cuda12.2 drivers, and then installs the Nvidia 470/Cuda11.4 drivers and fixes the version.

#!/bin/bash
yum remove -y cuda-toolkit* nvidia-driver-latest-dkms*
tmpfile=$(mktemp)
cat >$tmpfile <<EOF
[amzn2-nvidia]
name=Amazon Linux 2 Nvidia repository
mirrorlist=\$awsproto://\$amazonlinux.\$awsregion.\$awsdomain/\$releasever/amzn2-nvidia/latest/\$basearch/mirror.list
priority=20
gpgcheck=1
gpgkey=https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/7fa2af80.pub
enabled=1
exclude=libglvnd-*
EOF

mv $tmpfile /etc/yum.repos.d/amzn2-nvidia-tmp.repo
yum install -y system-release-nvidia cuda-toolkit-11-4 nvidia-driver-latest-dkms-470.182.03
yum install -y libnvidia-container-1.4.0 libnvidia-container-tools-1.4.0 nvidia-container-runtime-hook-1.4.0 docker-runtime-nvidia-1

echo "exclude=*nvidia* *cuda*" >> /etc/yum.conf
nvidia-smi

Create your own P2 compatible GPU-optimized AMI

You can create your own custom Amazon ECS GPU-optimized AMI that is compatible with P2 instances, and then launch P2 instances using the AMI.

  1. Run the following command to clone the amazon-ecs-ami repo.

    git clone https://github.com/aws/amazon-ecs-ami

  2. Set the required Amazon ECS agent and source Amazon Linux AMI versions in release.auto.pkrvars.hcl or overrides.auto.pkrvars.hcl.

  3. Run the following command to build a private P2 compatible EC2 AMI.

    Replace region with the Region with the instance Region .

    REGION=region make al2keplergpu

  4. Use the AMI with the following instance user data to connect to the Amazon ECS cluster.

    Replace cluster-name with the name of your cluster.

    #!/bin/bash
echo "ECS_CLUSTER=cluster-name" >> /etc/ecs/ecs.config