Using Trn1 and Inf1 instances on Amazon Linux 2 on Amazon ECS - Amazon Elastic Container Service
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China.

Using Trn1 and Inf1 instances on Amazon Linux 2 on Amazon ECS

You can register Amazon EC2 Trn1 and Amazon EC2 Inf1 instances to your clusters for machine learning workloads.

Amazon EC2 Trn1 instances are powered by Amazon Trainium chips, which are custom built by Amazon Web Services. These instances provide high performance and low cost training for machine learning in the cloud. You can train a machine learning inference model using a machine learning framework with Amazon Neuron on a Trn1 instance. Then, you can run the model on a Inf1 instance to use the acceleration of the Amazon Inferentia chips.

The Amazon EC2 Inf1 instances are powered by Amazon Inferentia chips, which are custom built by Amazon Web Services. They provide high performance and lowest cost inference in the cloud.

Machine learning models are deployed to containers using Amazon Neuron, which is a specialized SDK. The SDK consists of a compiler, runtime, and profiling tools that optimize the machine learning performance of Amazon machine learning chips. Amazon Neuron supports popular machine learning frameworks such as TensorFlow, PyTorch, and Apache MXNet.

Considerations

Before you begin deploying Neuron on Amazon ECS, consider the following:

  • Your clusters can contain a mix of Trn1, Inf1, and other instances.

  • You need a Linux application in a container that uses a machine learning framework that supports Amazon Neuron.

    Important

    Applications that use other frameworks might not have improved performance on Trn1 and Inf1 instances.

  • Only one inference or inference-training task can run on each Amazon Trainium or Amazon Inferentia chip. Each chip has two cores that are associated with it. You can run as many tasks as there are chips for each of your Trn1 and Inf1 instances.

  • When creating a service or running a standalone task, you can use instance type attributes when you configure task placement constraints. This ensures that the task is launched on the container instance that you specify. Doing so can help you optimize overall resource utilization and ensure that tasks for inference workloads are on your Trn1 or Inf1 instances. For more information, see Amazon ECS task placement.

    In the following example, a task is run on an Inf1.xlarge instance on your default cluster.

    aws ecs run-task \ --cluster default \ --task-definition ecs-inference-task-def \ --placement-constraints type=memberOf,expression="attribute:ecs.instance-type == Inf1.xlarge"
  • Neuron resource requirements can't be defined in a task definition. Instead, you configure a container to use specific Amazon Trainium or Amazon Inferentia chips available on the host container instance. Do this by using the linuxParameters parameter and specifying the device details. For more information, see Task definition requirements.

Using the Amazon ECS optimized Amazon Linux 2 (Inferentia) AMI

Amazon ECS provides an Amazon ECS optimized AMI that's based on Amazon Linux 2 for Amazon Trainium and Amazon Inferentia workloads. It comes with the Amazon Neuron drivers and runtime for Docker. This AMI makes running machine learning inference workloads easier on Amazon ECS.

We recommend using the Amazon ECS optimized Amazon Linux 2 (Inferentia) AMI when launching your Amazon EC2 Trn1 and Inf1 instances. You can retrieve the current Amazon ECS optimized Amazon Linux 2 (Inferentia) AMI using the Amazon CLI with the following command.

aws ssm get-parameters --names /aws/service/ecs/optimized-ami/amazon-linux-2/inf/recommended

The following table provides a link to retrieve the current Amazon ECS optimized Amazon Linux 2 (Inferentia) AMI IDs by Region.

Task definition requirements

To deploy Neuron on Amazon ECS, your task definition must contain the container definition for a pre-built container serving the inference model for TensorFlow. It's provided by Amazon Deep Learning Containers. This container contains the Amazon Neuron runtime and the TensorFlow Serving application. At startup, this container fetches your model from Amazon S3, launches Neuron TensorFlow Serving with the saved model, and waits for prediction requests. In the following example, the container image has TensorFlow 1.15 and Ubuntu 18.04. A complete list of pre-built Deep Learning Containers optimized for Neuron is maintained on GitHub. For more information, see Neuron Inference Containers.

763104351884.dkr.ecr.us-east-1.amazonaws.com/tensorflow-inference-neuron:1.15.4-neuron-py37-ubuntu18.04

Alternatively, you can build your own Neuron sidecar container image. For more information, see Tutorial: Neuron TensorFlow Serving on GitHub.

The task definition must be specific to a single instance type. You must configure a container to use specific Amazon Trainium or Amazon Inferentia devices that are available on the host container instance. You can do so using the linuxParameters parameter. The following table details the chips that are specific to each instance type.

Instance Type vCPUs RAM (GiB) Amazon ML accelerator chips Device Paths
trn1.2xlarge 8 32 1 /dev/neuron0
trn1.32xlarge 128 512 16 /dev/neuron0, /dev/neuron1, /dev/neuron2, /dev/neuron3, /dev/neuron4, /dev/neuron5, /dev/neuron6, /dev/neuron7, /dev/neuron8, /dev/neuron9, /dev/neuron10, /dev/neuron11, /dev/neuron12, /dev/neuron13, /dev/neuron14, /dev/neuron15
inf1.xlarge 4 8 1 /dev/neuron0
inf1.2xlarge 8 16 1 /dev/neuron0
inf1.6xlarge 24 48 4 /dev/neuron0, /dev/neuron1, /dev/neuron2, /dev/neuron3
inf1.24xlarge 96 192 16 /dev/neuron0, /dev/neuron1, /dev/neuron2, /dev/neuron3, /dev/neuron4, /dev/neuron5, /dev/neuron6, /dev/neuron7, /dev/neuron8, /dev/neuron9, /dev/neuron10, /dev/neuron11, /dev/neuron12, /dev/neuron13, /dev/neuron14, /dev/neuron15

The following is an example Linux task definition for inf1.xlarge, displaying the syntax to use.

{ "family": "ecs-neuron", "requiresCompatibilities": ["EC2"], "placementConstraints": [ { "type": "memberOf", "expression": "attribute:ecs.os-type == linux" }, { "type": "memberOf", "expression": "attribute:ecs.instance-type == inf1.xlarge" } ], "executionRoleArn": "${YOUR_EXECUTION_ROLE}", "containerDefinitions": [ { "entryPoint": [ "/usr/local/bin/entrypoint.sh", "--port=8500", "--rest_api_port=9000", "--model_name=resnet50_neuron", "--model_base_path=s3://your-bucket-of-models/resnet50_neuron/" ], "portMappings": [ { "hostPort": 8500, "protocol": "tcp", "containerPort": 8500 }, { "hostPort": 8501, "protocol": "tcp", "containerPort": 8501 }, { "hostPort": 0, "protocol": "tcp", "containerPort": 80 } ], "linuxParameters": { "devices": [ { "containerPath": "/dev/neuron0", "hostPath": "/dev/neuron0", "permissions": [ "read", "write" ] } ], "capabilities": { "add": [ "IPC_LOCK" ] } }, "cpu": 0, "memoryReservation": 1000, "image": "763104351884.dkr.ecr.us-east-1.amazonaws.com/tensorflow-inference-neuron:1.15.4-neuron-py37-ubuntu18.04", "essential": true, "name": "resnet50" } ] }