Give SageMaker AI Training Jobs Access to Resources in Your Amazon VPC

Configure a Training Job for Amazon VPC Access

To control access to your training jobs, run them in an Amazon VPC with private subnets that don’t have internet access.

You configure the training job to run in the VPC by specifying its subnets and security group IDs. You don’t need to specify the subnet for the container of the training job. Amazon SageMaker AI automatically pulls the training container image from Amazon ECR.

When you create a training job, you can specify the subnets and security groups in your VPC using the Amazon SageMaker AI console or the API.

To use the API, you specify the subnets and security group IDs in the VpcConfig parameter of the CreateTrainingJob operation. SageMaker AI uses the subnet and security group details to create the network interfaces and attaches them to the training containers. The network interfaces provide the training containers with a network connection within your VPC. This allows the training job to connect to resources that exist in your VPC.

The following is an example of the VpcConfig parameter that you include in your call to the CreateTrainingJob operation:

VpcConfig: {
      "Subnets": [
          "subnet-0123456789abcdef0",
          "subnet-0123456789abcdef1",
          "subnet-0123456789abcdef2"
          ],
      "SecurityGroupIds": [
          "sg-0123456789abcdef0"
          ]
        }

Configure Your Private VPC for SageMaker AI Training

When configuring the private VPC for your SageMaker AI training jobs, use the following guidelines. For information about setting up a VPC, see Working with VPCs and Subnets in the Amazon VPC User Guide.

Ensure That Subnets Have Enough IP Addresses

Training instances that don't use an Elastic Fabric Adapter (EFA) should have at least 2 private IP addresses. Training instances that use an EFA should have at least 5 private IP addresses. For more information, see Multiple IP addresses in the Amazon EC2 User Guide.

Your VPC subnets should have at least two private IP addresses for each instance in a training job. For more information, see VPC and Subnet Sizing for IPv4 in the Amazon VPC User Guide.

Create an Amazon S3 VPC Endpoint

If you configure your VPC so that training containers don't have access to the internet, they can't connect to the Amazon S3 buckets that contain your training data unless you create a VPC endpoint that allows access. By creating a VPC endpoint, you allow your training containers to access the buckets where you store your data and model artifacts. We recommend that you also create a custom policy that allows only requests from your private VPC to access to your S3 buckets. For more information, see Endpoints for Amazon S3.

Use a Custom Endpoint Policy to Restrict Access to S3

The default endpoint policy allows full access to S3 for any user or service in your VPC. To further restrict access to S3, create a custom endpoint policy. For more information, see Using Endpoint Policies for Amazon S3. You can also use a bucket policy to restrict access to your S3 buckets to only traffic that comes from your Amazon VPC. For information, see Using Amazon S3 Bucket Policies.

Restrict Package Installation on the Training Container

The default endpoint policy allows users to install packages from the Amazon Linux and Amazon Linux 2 repositories on the training container. If you don't want users to install packages from that repository, create a custom endpoint policy that explicitly denies access to the Amazon Linux and Amazon Linux 2 repositories. The following is an example of a policy that denies access to these repositories:

{ 
    "Statement": [ 
      { 
        "Sid": "AmazonLinuxAMIRepositoryAccess",
        "Principal": "*",
        "Action": [ 
            "s3:GetObject" 
        ],
        "Effect": "Deny",
        "Resource": [
            "arn:aws:s3:::packages.*.amazonaws.com/*",
            "arn:aws:s3:::repo.*.amazonaws.com/*"
        ] 
      } 
    ] 
} 

{ 
    "Statement": [ 
        { "Sid": "AmazonLinux2AMIRepositoryAccess",
          "Principal": "*",
          "Action": [ 
              "s3:GetObject" 
              ],
          "Effect": "Deny",
          "Resource": [
              "arn:aws:s3:::amazonlinux.*.amazonaws.com/*" 
              ] 
         } 
    ] 
} 

Configure Route Tables

Use default DNS settings for your endpoint route table, so that standard Amazon S3 URLs (for example, http://s3-aws-region.amazonaws.com/amzn-s3-demo-bucket) resolve. If you don't use default DNS settings, ensure that the URLs that you use to specify the locations of the data in your training jobs resolve by configuring the endpoint route tables. For information about VPC endpoint route tables, see Routing for Gateway Endpoints in the Amazon VPC User Guide.

Configure the VPC Security Group

In distributed training, you must allow communication between the different containers in the same training job. To do that, configure a rule for your security group that allows inbound connections between members of the same security group. For EFA-enabled instances, ensure that both inbound and outbound connections allow all traffic from the same security group. For information, see Security Group Rules in the Amazon Virtual Private Cloud User Guide.

Connect to Resources Outside Your VPC

If you configure your VPC so that it doesn't have internet access, training jobs that use that VPC do not have access to resources outside your VPC. If your training job needs access to resources outside your VPC, provide access with one of the following options:

Monitor Amazon SageMaker Training Jobs with CloudWatch Logs and Metrics

Amazon SageMaker AI provides Amazon CloudWatch logs and metrics to monitor training jobs. CloudWatch provides CPU, GPU, memory, GPU memory, and disk metrics, and event logging. For more information about monitoring Amazon SageMaker training jobs, see Metrics for monitoring Amazon SageMaker AI with Amazon CloudWatch and SageMaker AI jobs and endpoint metrics.